Meta AI Pivots Strategy Again as Mark Zuckerberg Releases Muse Glimmer and 6,500-Word AI Manifesto

With new open models, Meta pitches another reboot of its struggling AI strategy

Meta has announced its intention to focus on open-weight large language models. Additionally, the company announced the release of an open model called Muse Glimmer and a promise to open the weights for Muse Spark 1.2, its more powerful model, in the next few weeks. Alongside these announcements, Meta CEO Mark Zuckerberg published a more than 6,000-word essay outlining the company's philosophy about AI systems and governance moving forward. The essay aims to differentiate Meta from companies like OpenAI and Anthropic, which develop proprietary models and which have lobbied the US government for help competing against large-scale distillation -- which involves using an existing model to train a new one -- or open-weight models by Chinese labs. Muse Glimmer is a 30 billion parameter model with a 128,000-token context window by default. It is distilled from Muse Spark, the larger and more capable model that Meta launched earlier this year. Glimmer is meant to run on users' local machines, rather than via a cloud service or an API. Glimmer's weights are open source under the Apache 2.0 license. Muse Spark was introduced in April as a closed, proprietary, frontier-class model -- Meta's first major model release after a significant shake-up of the company's AI teams last year, and a departure from its focus on models that are, by some definition, open. When Meta released Muse Spark 1.1 in July, it introduced its first paid service -- again, a departure from its previous strategy. Muse Spark 1.2 was released on August 5 and was accompanied by Muse Code, a terminal coding agent. Developers have generally found that Muse Code doesn't quite match the frontier models from Anthropic or OpenAI in capability, but it competes well on cost -- meaning it has similar positioning to many open-weight models from Chinese labs. As a smaller model designed to run on consumer GPUs, Muse Glimmer won't compete on that level at all -- but it reflects a growing movement to bring some inference to local devices to reduce reliance and spending on the models produced by the big labs like Anthropic and OpenAI. Open weights, open access Alongside the model releases, Mark Zuckerberg is credited as the author of a lengthy open letter that details Meta's corporate strategy with AI, and a broader argument for how AI should be developed, distributed, and regulated. It follows several statements by other Big Tech and AI company leaders debating the merits of open-weight models, proprietary labs, and the practice of distillation -- something that some Chinese labs have reportedly done to build models that compete with the latest efforts from Anthropic and others. On July 24, several companies including Nvidia, Hugging Face, Meta, Mistral, Mozilla, OpenAI and others co-signed an open letter titled "Open Weights and American AI Leadership" that argued for the value of open-weight models as opposed -- or at least in addition -- to proprietary ones, and defended distillation as a legitimate practice, even as it carved out a distinction for "unlawful efforts to extract value from closed models." It advocated for "targeted legal and commercial frameworks rather than sweeping restrictions on techniques that play an important role in AI innovation." On the topic of distillation, Meta and Zuckerberg wrote: The ability for models to learn from other models is an important principle of how the open source ecosystem works. All AI models are derived from human knowledge. Some have tried to frame distillation as harmful, but I think it is important to protect the principle that you can learn from anything you can observe. However, the majority of the essay focused on the notion of decentralized AI systems that are distributed widely. It also directly aims at arguments that tightly controlled, proprietary AI models and systems are needed because of concerns about existential threats or catastrophic misalignment: It is surprising that the discourse from many developing AI is so filled with doom. I do not understand why anyone who believes that AI will eliminate most jobs and much of humanity's relevance would rush to build that future. The notion that AI is so dangerous that the only safe path is an extreme concentration of power seems inherently problematic. Zuckerberg wrote that the view of alignment taken by companies like Anthropic -- which claims that it is trying to train singular and broad foundation models with operational parameters and guardrails that will ensure they benefit humanity broadly -- is "fundamentally flawed." "People's diverse values represent different tradeoffs they would make on important issues. There is no technological solution that can align with everyone's opposing interests and values at once," the essay says. "Any singular superintelligence would have to prioritize some values over others and in the process would be incapable of being benevolent to everyone." Instead, Meta argues here that models should be personalized to the needs and values of individuals or groups of individuals. It also claims that decentralization will make everyone safer, because it will give the benefits and advantages of "superintelligence" to everyone equally, instead of privileging "a small number of individuals, businesses, governments, or AI itself." A public recalibration Meta has been lagging behind other big tech companies and major frontier labs for foundation models. Its models haven't seen the kind of adoption that those developed by OpenAI or Anthropic have. OpenAI and Anthropic have aggressively targeted enterprise customers, releasing powerful models and harnesses for knowledge work tasks like software development, and they have made significant inroads and generated substantial revenue from this strategy. Meta has not seen the same level of success. Meta also saw a total overhaul of its AI division last year, when former Meta AI chief scientist Yann LeCun was replaced by former Scale AI CEO Alexandr Wang. The reset led to a change in focus. In recent months, the debate around open-weight models and distillation has increased in volume as recent Chinese models like Alibaba's Qwen3.8-Max and Moonshot's Kimi K3 have been shown to rival Anthropic and OpenAI at the frontier. Those models may perform slightly worse in coding benchmarks, for example, but they are generally cheaper to use. In a sense, Meta seems to be positioning itself as a US alternative to Alibaba, Moonshot, or DeepSeek -- not quite as frontier-facing as Anthropic or OpenAI, but more open, customizable, and affordable. It is also orienting itself -- at least with these public statements -- more toward personal use as opposed to large-scale enterprise deployments, at least for now. That is a retreat from some of its earlier ambitions, in a way, as the company takes advantage of changing winds to try to plot a new course.

Mark Zuckerberg's AI manifesto is exactly why people don't like AI

On Monday, Mark Zuckerberg published a 6,500 word manifesto about personal AI, largely about the possibilities for the "personal superintelligence" systems Meta AI is building. The ideas in the post aren't totally new. A version of the essay ran in the Wall Street Journal a few months ago and he's talked about them on Meta earnings calls before. But this is probably the most detailed version he's shared. I've seen it described as an anti-doomer essay, but that's not quite right. It's more of a description of why Mark Zuckerberg is personally excited about the ways AI is going to change society. Yet, even as Zuckerberg tries to paint a picture of the wonderful future abundant superintelligence will bring, he keeps reminding us of all the ways it's likely to go wrong. I find AI exciting too -- that's why I keep writing about it -- but a lot of the public sees AI as creepy and unpleasant. So it's worth pinning down exactly what's happening here, and why Zuckerberg isn't doing the industry any favors with essays like this. Social media is still a sensitive topic in the tech industry, and we don't have the space here to litigate the relative merits of every single complaint people have about Facebook. Suffice it to say, Facebook as a prodcut and Zuckerberg as a person are both unpopular with the US public. A recent survey found that 64 percent of Americans believe social media has been harmful to democracy and a similar percentage believe it should be more heavily regulated, numbers that cut evenly across partisan lines. Just this weekend, a court fined the company $567 million for being harmful to children. The vibes are bad. I don't bring this up to imply that Zuckerberg should withdraw to the wilderness in shame -- but the fallout from social media is one of the central reasons we're now seeing so much anxiety about the social impact of AI. Whether it's fair or not, the public does not trust tech executives to make sure new technologies like this have a positive impact on society. Instead of acknowledging that and trying to win back their trust, this essay demonstrates over and over again how the trust was lost in the first place. A large part of the essay is devoted to hazy generalities about intelligence -- very similar to the hazy generalities Zuckerberg and Dorsey used to give about free speech. Here is one for instance: As everyone gains more powerful tools, each person will become more capable of shaping the future, not less....People and institutions with competing interests naturally check and balance each other to lead towards positive outcomes. The best and most realistic path to building a positive AI future is by delivering superintelligence to everyone. I guess? I can think of a few examples of conflicting interests leading to bad outcomes, but let's put that aside. The weirder thing is that this is being presented as an earnest philosophical conclusion, instead of a specific product called "personal intelligence" that Meta is bringing to market. If you're already inclined to distrust this person, you might worry that they're writing all this to convince themselves that nothing bad can happen. That feeling got more intense for me when Zuckerberg got to the specific examples, many of which are alarmingly out of touch with the reality of how AI tools are being used. For instance, this is Zuckerberg's take on AI in education: Everyone will have a personalized tutor and coach with a PhD in every subject and unlimited patience to help you learn anything you want. Students will have extra help in areas they need it that is currently only available to those whose parents can pay. Adults will have a superintelligent learning assistant that knows exactly how to teach you new job skills, new languages, new hobbies, or anything else you're interested in. Of course, the product Zuckerberg is describing already exists: this is a consumer chatbot like ChatGPT, Claude and Gemini. These tools really are helpful if you want to learn about a certain topic, but the main way they're used in education is to avoid learning, since your personalized tutor can do your homework and write your assigned essay -- and because there's no robust watermarking system in place, there's no way for teachers to be sure which essays were generated with AI. This is a pretty low-stakes example of AI harms, but it's a real thing that is happening right now. Sometimes you design a technology to do good things and it has unintended consequences that make things worse. And the fact that one of the most powerful people in the world refuses to acknowledge can make people understandably nervous! Zuckerberg also gives an example of how AI will change the legal system for the better: As a thought experiment, imagine only one person had a superintelligent lawyer. They would have an unfair advantage in court -- even if they were wrong on the merits. That would lead to a worse society. But now imagine everyone has a superintelligent lawyer. In this case, justice would be carried out much more fairly and efficiently than it is today when there is often an imbalance in skills and resources in litigation. Again, that's another loaded example. Sure, access to legal AI might lead to a more just society, but it also might add more complexity to the bureaucratic hellscape that already exists. Or it might unleash a new wave of vexatious litigants that just clog up the works with the legal equivalent of spam. It's hard to feel calm about any of this stuff, and the fact that Zuckerberg isn't worried makes me more worried. In other places, even descriptions of Meta's existing business practices start to take on an oddly abstract character. For instance, this is how Zuckerberg describes Meta's commitment to a freemium model. Everyone will have free or affordable access to these tools.... We will offer free versions that will be accessible to billions of people. For those who want to pay to use more compute, there will be a dynamic auction mechanism that will guarantee that everyone gets the lowest price possible for the intelligence and compute they're using while also ensuring the capacity is used for whatever people collectively find most valuable. This will ensure the benefits of superintelligence are distributed widely. In broad strokes, I agree with everything he's saying here. Zuckerberg is right that access is an issue, and he's right that a freemium model allows for more access than requiring everyone to pay for compute up front. Dynamic markets for spot compute already exist, so it's not like he's describing something completely outside the norm. But there's a reason every consumer AI product insulates end-users from the spot price of the compute they're using. Surge pricing is a terrible user experience, particularly when it's deployed on a tool you actually rely on for your work. Again, I have to assume that Zuckerberg knows this, and he's not actually planning to set token prices through dynamic auctions. But I'm actually less sure of that now than I was before I read this piece. There are better ways to talk about this stuff. For all their faults, Sam Altman and Dario Amodei do a pretty good job when they're directly communicating with the public. They do exactly what Zuckerberg resists here: acknowledging the dangers of AI, emphasizing their own precautions, and trying to convince listeners that they can be trusted. That communication can only do so much. Sometimes (especially recently), the safeguards fail and the inherent danger of the whole enterprise is made clear. But without some kind of trust-building, the industry may not always be able to survive those failures.

Meta's new Glimmer AI model offers a hint at Zuckerberg's personal intelligence vision

Meta on Monday released Muse Glimmer, an open-weight model designed to power AI agents locally on consumer hardware, providing the clearest picture yet of what CEO Mark Zuckerberg's vision of "personal superintelligence" could look like in practice. The 30-billion parameter model is essentially an open version of Meta's most powerful closed model, Muse Spark, which the company debuted in April. Glimmer's weights are available under the permissive Apache 2.0 license, so developers can download them and modify as necessary. Glimmer is designed to run AI agents that can perform multi-step tasks -- like call tools, write and debug code, work with files and screenshots, and execute on a task over an extended workflow -- locally on a Mac or PC with a single consumer GPU. It supports text and images, and was trained across more than 100 languages, the company said. Meta imagines Glimmer being used for things like managing schedules, drafting messages, and organizing files -- tasks that would require large amounts of access to personal data. By processing the information on a user's device instead of sending it to the cloud, Meta is laying the groundwork for a more privacy-sensitive personal agent. Glimmer is also designed to be "always-on" and able to operate "anywhere, anytime, with or without an internet connection." That vision mirrors the future Mark Zuckerberg has previously laid out for Meta. Last year, he argued that advanced AI should empower individuals rather than stay concentrated in the hands of a few companies, while also warning that Meta would have to be careful about which of its increasingly powerful models it released openly due to safety concerns. In a new letter Monday, Zuckerberg reiterated that vision, arguing that distributing superintelligence widely "has the potential to begin a new era of personal empowerment where individuals can use this powerful new capability to reach their full potential, pursue their interests, and improve their lives and the world more than ever before." Zuckerberg went on to list the ways Meta's superintelligence can improve a person's life, from providing a capable personal agent that "will work 24/7 on your behalf to improve your relationships, health, career, finances, home management, hobbies, and more" to giving people access to tools needed to create a new business or advance scientific progress. The biggest promise of all is that "everyone will have free or affordable access to these tools." But access isn't the same as ownership. Zuckerberg's promise to distribute superintelligence widely comes as Meta is increasingly distinguishing between models it will release openly and those it will keep under its control. Muse Spark, its more powerful mode, remains closed-weight, while the smaller Glimmer can be downloaded, fine-tuned, and run on a user's hardware. As a result, Glimmer offers an early indication of where Meta may draw the line between the AI it wants people to own themselves and the more powerful intelligence that remains under the company's control.

Meta's New Open-Weight Model Can Run AI Agents on Your Laptop - CNET

Katelyn is a reporter with CNET covering artificial intelligence, including chatbots, image and video generators.... Read full bio Meta is putting its money (and AI models) where its mouth is. The company said Monday that its latest model is built specifically for AI agents. Meta Glimmer is out now and is built to run on your laptop or PC. Meta Muse Glimmer is a 30-billion-parameter model, which sounds big, but it's actually built to run more efficiently than other big AI models. (By comparison, Moonshot AI's Kimi K3 model, which approaches the capabilities of top models from companies like Anthropic and OpenAI, has 2.8 trillion parameters.) It can run on a single GPU, so you can run it on a newer MacBook or another computer with a relatively recent chip. The model is meant to help you orchestrate an army of AI agents, the newest wave of generative AI tech that lets you outsource tasks to bots known as agents. Open-weight AI lets anyone, including researchers and cybersecurity defenders, see how an AI model operates. It isn't fully transparent - open-source advocates say that to be truly open source, we'd need to know what's inside the black box of training data used by Meta and other companies - but it's less opaque than the closed-weight models offered by other major US developers. You can check out the model's documentation on Meta's website and HuggingFace. Meta CEO Mark Zuckerberg has been one of the biggest proponents of open-source, or open-weight, AI. He writes in an essay published Monday that the way to secure our future is to ensure that access isn't limited to a few institutions, including tech companies like OpenAI, Anthropic and Meta. (It's worth noting that Zuckerberg doesn't have the same open philosophy when it comes to his social media platforms, which are run by black-box algorithms.) "The defining questions of our age are who will have access to superintelligence and what will we direct it towards," Zuckerberg wrote. "We propose a philosophy based on individual empowerment as the source of prosperity, invention as the primary purpose of superintelligence, and balance of power as the foundation of safety." The company is also introducing a $1 billion fund to "support each community" where it opens data centers. Data centers are the controversial facilities tech companies need to keep AI running, but towns across the US have fought hard to keep them out due to environmental concerns, noise, higher electricity prices and more. New York recently became the first state to pass a moratorium on data center construction. A billion dollars is a drop in the bucket for the $1.5 trillion company, but Meta says its investment in the area can include giving teachers bonuses, funded by tax revenue. Meta spent billions of dollars over the last few years to staff up its Superintelligence Labs with some of the biggest names in the AI field, and it seems the group is now finally showing progress. The company has released a slew of new models recently, including the closed-weight models Meta Spark 1.1 and Spark 1.2, which power its Meta Code tool that competes with Anthropic's Claude Code and OpenAI's Codex. It also dropped an AI image tool that let anyone create deepfakes on Instagram, which was pulled less than a week later amid intense public backlash. Meta is one of the only big tech companies to release open-weight AI models. OpenAI released a family of them in 2025, called GPT-OSS. Anthropic's Claude models are all closed, which means we have little insight into how they operate.

Four takeaways from Mark Zuckerberg's massive AI manifesto

Meta CEO Mark Zuckerberg has a lot to say about the idealized future he now envisions for humanity co-existing with artificial intelligence -- his latest essay spans more than 6,500 words on the matter. The lengthy manifesto Zuckerberg published on Monday, titled "The Future is for Everyone," broadly lays out his beliefs about how the technology should be developed, expanded, and regulated, and how Meta is positioning itself to enable those goals. Some points echo a similar (albeit, much shorter) public letter that Zuckerberg put out last year, heralding the importance of public access to superintelligent AI -- a term for artificial general intelligence (AGI) models capable of surpassing human intelligence, which AI companies are racing to develop. Both were published with similar intentions: to pitch the supposed personal, societal, and economic benefits of superintelligent AI, and to grandstand Meta as a responsible steward to deliver that future. Here are some of the main takeaways from Zuckerberg's latest essay: "Sustainable infrastructure development means that communities must benefit significantly from each project," Zuckerberg said, detailing how Meta plans to resolve the ongoing PR crisis surrounding AI datacenter projects. This includes pledges to build energy-generating infrastructure "wherever we invest" that could potentially give low-cost energy back to the communities, and a commitment to "restore more water than we use in the watersheds where we operate by 2030." Meta is also launching a Future Is For Everyone Fund to support local communities, and Zuckerberg says the company is providing free training and "guaranteed high-paying jobs" for skilled tradespeople in areas where Meta is building data centers. Meta is planning to focus its efforts on "delivering personal superintelligence to billions of people and small businesses," arguing that doing so will empower everyone to become more entrepreneurial. Part of this involves offering the technology "for free or as affordably as possible," and leaning into releasing more open source AI models that anyone can download. Doing so "will bring an abundance of personal agency, expression of ideas, deepening of relationships, economic and financial prosperity, improved health, and inventions we cannot imagine today," according to Zuckerberg. He also says this will have an impact on "some aspects of the way we work," but waves off concerns that AI will result in fewer jobs overall. Zuckerberg voices the importance of the US leading the open source AI ecosystem, citing concerns that competing models released by foreign labs "hold several advantages" because they face fewer training data restrictions. He pitches that the US needs to rethink policies around distillation and data use, arguing that while "some have tried to frame distillation as harmful," the US will not be able to keep up if it keeps restricting AI models from learning from other models. "I do not believe restricting access to foreign open source models is an effective solution," said Zuckerberg. We should note here that these concerns are likely being voiced in response to the release of open-weight models from Chinese AI companies like Alibaba and Moonshot, which isn't the same thing as being open-source, and the models Meta has released so far aren't truly open-source either. He also espouses that access to personal superintelligence is essential for individuals to maintain their freedom, and should only be subject to restrictions "when truly required." In cases where cybersecurity is involved, Zuckerberg proposes that frontier AI labs should work with the US government to help with bolstering critical infrastructure, and believes open-source AI models will improve system security over time -- especially "once superintelligence enables most of the world's code to be verifiably secure." That plan includes pushing AI developers to share the training information of new models "for government use and review" intermediately instead of waiting until training has completed, which aligns with the Trump administration's AI testing framework. "This way the government will have early access to the most powerful models and an army of capable engineers to identify and patch security issues," Zuckerberg said. "Labs should also work with law enforcement to help identify bad actors attempting to misuse their systems.

Zuck rekindles open weights Llama drama with Muse Glimmer

After seemingly abandoning its open source AI roots this spring, Meta offered enterprises a glimmer of hope on Monday with the launch of its first open weights model in more than a year. Unveiled on Monday, Muse Glimmer is a 30 billion-parameter LLM distilled from the Social network's larger, and for now proprietary, Muse Spark model. Glimmer arrives as American tech companies grapple with a crisis over the proliferation of Chinese open weights models and a call for domestically-created alternatives. Meta built its reputation on the back of open weights model development beginning with the Llama herd back in 2023. But after Llama 4 flopped and the company restructured its AI group, critics questioned CEO Mark Zuckerberg's commitment to open source AI. With Muse Glimmer, Meta has returned to the open weights arena. The company describes the model as being ideally suited to local AI inference workloads, including local agents, code assistants, and applications requiring robust multi-modal tool use and function calling. Released under a highly permissive Apache 2.0 license, enterprises are also free to deploy, use, and modify the model however they see fit. We imagine it won't be long before Nous Research -- one of the original Llama fine-tuners -- emits another Hermes model based on Glimmer. Early support has begun to hit popular local AI inference platforms like Llama.cpp, Ollama, and Unsloth, with optimized implementations expected to hit over the next few weeks. But at 30 billion parameters, Muse Glimmer doesn't exactly move the needle much on reclaiming American open weights superiority. It's too small to compete with Moonshot AI's Kimi K3, Alibaba's Qwen 3.8-Max, DeepSeek V4 Flash, or any of the other Chinese models which have dominated the AI news cycle over the past few months. Instead, Muse Glimmer is positioned as a model for small-to-medium sized enterprises or enthusiasts, competing with similarly-sized LLMs from Alibaba and Google. This is reflected in Meta's benchmark figures, which pit the model against Alibaba's Qwen 3.6-27B and Google Gemma 4 31B. As usual, take these claims with a grain of salt, but Glimmer does appear to best Google's Gemma in most scenarios, and trades blows with Alibaba's equivalently sized model. Unfortunately for Meta, the comparison probably won't age well, with Qwen 3.8-27B due to be released any day now. Open and local Glimmer's relatively small parameter count means that its hardware requirements are rather modest compared to larger frontier-class models, like DeepSeek V4. At its native BF16 precision, the model should fit comfortably into a single Nvidia RTX Pro 6000 or AMD MI350P. Quantized to 4-bit precision and model's weights shrink from around 60 GB to just under 16 GB -- small enough to fit in a 20 to 24 GB consumer graphics card, like an RTX 30/4090 or RX 7900 XT/XTX. Unfortunately for those with 16 GB cards, a lack of adequate working memory means dropping down to a smaller, less accurate 3-bit quant. Even if you can get the model to fit, it won't be as fast as some other recent models, like Qwen 3.6-35B-A3B or Gemma 4 26B-A4B, as Glimmer uses all 30 billion parameters to generate each token where the others use just 3 to 4 billion. Glimmer benefits heavily from memory bandwidth. On cards like the RTX 5090, which has 1.8 TB/s of it, Meta says users can expect between 75 and 233 tok/s. The higher end of that relies on a technique called speculative decoding. As a quick refresher, speculative decoding uses a small draft model, in this case one modeled after DeepSeek's DSpark drafter, to speed up inference by predicting the outputs of a larger model. If you're interested, we explore the concept in greater depth in this hands-on here. While 233 tok/s is plenty fast for most agentic workloads, most users won't see performance nearly that high. On an M5 Max MacBook Pro, Meta estimates that the systems' up to 614 GB/s of bandwidth will deliver a still-meaningful 26.2 to 57.8 tok/s, but it's also worth pointing out the M5 Max delivers 4x the memory bandwidth of the typical Windows notebook today. In other words, unless you've got a dedicated graphics card with enough memory to run the model, the best you can expect is around 6 to 14 tok/s. Testing on a DGX Spark in Unsloth Studio, we were getting around 12.2 tok/s, though DSpark support doesn't appear to have been implemented just yet. If you'd like to try the model out for yourself, its weights are currently available for download on Hugging Face as well as through local inference platforms like Ollama and LM Studio. Sparking joy Meta's return to the open weights arena doesn't stop with Glimmer. In an X post Monday, Meta's Superintelligence chief Alexandr Wang committed to releasing an open weights version of Muse Spark 1.2 "soon." Muse Spark 1.2 is Meta's most capable model, its first to compete in a meaningful way with rivals OpenAI, Anthropic, and Google. However, even this release won't close the gap with Chinese models. Both Kimi K3 and Qwen 3.8-Max outperform Spark, according to Artificial Analysis' intelligence index. Having said that, we still don't know how big Spark actually is. If Meta's next open weights model ends up delivering similar performance using fewer weights, this may end up being more competitive from a cost per token standpoint. ®

Meta launches new AI model as Zuckerberg champions open-weight push

Aug 10 (Reuters) - Meta (META.O), opens new tab CEO Mark Zuckerberg called for lower U.S. barriers for open-source AI models to compete with Chinese rivals as the social media giant released a new open-weight model on Monday. The new model, Muse Glimmer, is much smaller than leading AI models from rivals and is instead designed for agentic tasks and can run on a Mac or PC with a single graphics card, aiming to tap demand for AI systems that run directly on people's devices. It comes as the social media giant seeks to strengthen its position after forming a costly, new superintelligence team last year to propel itself back into the high-stakes AI race. Zuckerberg's statement also marks the latest show of support for open-weight AI, which is gaining traction ⁠as businesses grow wary of ballooning AI bills and worry about recent cybersecurity incidents involving models from Anthropic, OpenAI and Meta. Hugging Face, the AI coding collaboration site that was hacked by a rogue OpenAI model, said last month that it used a Chinese open-weight model to defend against the attack because closed-source models have restrictions on use for cybersecurity work. Open-weight models are typically cheaper than leading models from so-called frontier labs such as OpenAI and Anthropic. Open-weight models also come with publicly accessible core components for easy customization, unlike closed models that companies keep fully under their control. POLICY RETHINK NEEDED TO PROPEL OPEN-WEIGHT Zuckerberg said in a statement that the U.S. needed to rethink policies if domestic ⁠firms were to lead in open-weight models. Chinese startups are leading the race for open-weight models, with Moonshot's Kimi K3, alongside Alibaba's Qwen3.8-Max and DeepSeek's V4-Flash, delivering performance that rivals top systems by U.S. AI labs. By contrast, the leading models of U.S. developers OpenAI, Anthropic and Alphabet's (GOOGL.O), opens new tab Google are closed source. "Foreign labs currently hold several advantages here since American labs have to ⁠comply with many additional restrictions on training data," Zuckerberg said, referring to open-source models. "US policy must reduce this additional friction if we want American open source models to lead over time," Zuckerberg said, adding that restricting access to foreign open-source models was not an ⁠effective solution. U.S. President Donald Trump's administration told AI developers earlier this month that it will not put open-weight AI models through voluntary safety tests, according to two sources familiar with the discussions. In his statement, Zuckerberg also advocated for AI ⁠model distillation, or using a powerful AI system to train a smaller model. He said that Meta would implement a governance structure to give its independent directors the power to approve the safety criteria for releasing models. Reporting by Ananya Palyekar and Shubham Kalia in Bengaluru; Editing by Mrigank Dhaniwala and Devika Syamnath Our Standards: The Thomson Reuters Trust Principles., opens new tab

Mark Zuckerberg attacks 'closed' AI rivals as Meta returns to open models

Meta's founder casts OpenAI and Anthropic as foils in his pitch for powerful AI to become more freely available Meta chief Mark Zuckerberg has argued against allowing powerful AI to become concentrated in the hands of companies and governments, as the Big Tech giant resumes making some of its AI models freely available to outside developers. The $1.5tn company on Monday released the underlying parameters of Muse Glimmer, a new "open" AI model that developers will be able to download and modify. Meta said it would also release those parameters, known as "weights", for a version of its more powerful Muse Spark model in the coming weeks. Alongside the announcement, Zuckerberg published an essay on Meta's website outlining a plan to distribute "free" and highly capable AI to "billions of people", in an effort to empower individuals and "check and balance the power of institutions". In an apparent attack on groups such as Google, Anthropic and OpenAI, he wrote: "Most other labs are focused on building AI for companies, governments, or other institutions. So if those labs lead, then the balance of power will favor larger institutions over individuals". The pledges marks a return to the "open" AI strategy that Meta has used to distinguish itself from rivals. Meta declined to release the underlying weights of Muse Spark earlier this year, citing safety concerns, as competition among leading AI companies increasingly focuses on how widely their most capable technology should be distributed. On Monday, Meta also unveiled a $1bn fund for communities hosting its US data centres as it accelerates the infrastructure build-out needed to support those ambitions. Big Tech companies have been the target of a backlash from local residents near data centres, as the developments can compete for scarce resources such as power and water. Zuckerberg's intervention comes as Meta spends billions of dollars chasing the frontier of AI and promoting his vision of "personal superintelligence" -- advanced AI agents intended to help individuals with everything from health and hobbies to finances and careers. But Meta's share price fell nearly 8 per cent after it revealed last month that its spending on AI infrastructure had pushed its free cash flow down by 91 per cent. The social media group's shares have fallen nearly a fifth over the past year, as it has poured money into hiring top AI talent and building AI data centres. Zuckerberg suggested the company would press on with investing in open-weight models, criticising the "closed" approach of rival labs that have made billions in revenue from selling access to their models. "Some argue that the best way to reduce risk is to restrict the capabilities individuals can access," wrote Zuckerberg, arguing that "widely deployed open source systems have proven more secure because more people can identify vulnerabilities". He gave the example of Hugging Face, a start-up that relied on open models to patch its security systems after closed models refused its requests on safety grounds. Zuckerberg's comments came as Chinese AI labs narrow the gap between their own open models and the most advanced proprietary offerings from US labs. OpenAI and Anthropic have accused Chinese rivals of training their models through the outputs of US labs, using a technique called distillation. However, Zuckerberg came out in favour of the practice, saying it was "important to protect the principle that you can learn from anything you observe" and rejecting the idea that distillation was "harmful". Zuckerberg said that Meta would place the power to review the safety of its models in the hands of its independent board of directors. The Meta founder criticised rival AI labs for espousing a "discourse . . . so filled with doom". "The notion that AI is so dangerous that the only safe path is an extreme concentration of power seems inherently problematic," he said. "As everyone gains more powerful tools, each person will become more capable of shaping the future, not less."

Meta's latest model advances Zuckerberg's vision for personal AI assistants

Every weekday, the CNBC Investing Club with Jim Cramer releases the Homestretch -- an actionable afternoon update, just in time for the last hour of trading on Wall Street. Markets are off to a mixed start to the week. The S & P 500 traded higher for most of the early afternoon, but the bulls struggled to keep the ball in the air. What held back the broader market's attempted rally was oil. WTI crude climbed 4% as markets awaited more clarity on a potential deal with Iran to open the Strait of Hormuz. Also not helping the bull's cause was data from the Department of Energy that showed the U.S. Strategic Petroleum Reserve has fallen below 300 million barrels for the first time since the 1980s. The rise in energy prices pressured bonds, pushing interest rates higher and sending the 10-year Treasury yield back to 4.70% ahead of two key inflation reports later this week. The odds of a September rate hike fell last Friday after the weak jobs report, and the market now thinks it's a coin flip between the Fed hiking rates or leaving them unchanged at its next meeting, according to the CME FedWatch tool. Meta unveiled the Muse Glimmer AI model on Monday, giving investors a glimpse into CEO Mark Zuckerberg's plans to bring artificial intelligence directly to consumers. It's also the latest release from its Meta Superintelligence Labs, the social media giant's high-dollar AI division led by Alexandr Wang. Unlike many AI systems that rely on cloud computing, Muse Glimmer is designed to run locally on personal devices including Macs and personal computers. Meta said the open-weight model is capable of agentic tasks, meaning it can perform actions and complete requests on behalf of users. An open-weight model is one whose code is publicly available for anyone to copy or modify. Meta made the model's weights publicly available through Hugging Face. "Running models locally enables you to use AI anywhere, anytime, with or without an internet connection," Meta said in a blog post. The company said it's working with partners like fellow Club names Nvidia and Intel , as well as Arm Holdings , Advanced Micro Devices , and Dell , to optimize the model's performance across a range of devices. It's worth remembering that Nvidia in June announced new PC chips for a lineup of Windows PCs set to debut later this year -- precisely the kind of devices that might run an AI model like Muse Glimmer locally. Intel is a longtime juggernaut in the PC market. The launch of Muse Glimmer coincided with the release of a lengthy letter from Zuckerberg , with the CEO outlining his vision for what he called a "positive AI future" that creates a "balance of power that favors individuals." Zuckerberg argued that AI should become deeply personal, helping people accomplish daily tasks and learn news skills to navigate everyday life. Muse Glimmer fits with that strategy, positioning the AI model as a personal assistant that runs locally on computers rather than a tool accessed only through the cloud. By putting powerful AI tools into users hands, it also gives them another reason to stay inside Meta's ecosystem, from Facebook, Instagram and WhatsApp to their on-device model. In the Monday letter, Zuckerberg also called on the U.S. and its allies to "lead the open source AI ecosystem that will make up a large percent of global AI use." Muse Glimmer checks the box there. Additionally, in an Instagram video on Monday, Zuckerberg said in the coming weeks, the company plans to open the weights of Muse Spark 1.2, its latest foundation model. The Muse Spark model family was first released in April and has been closed weights since then, so Zuckerberg's update is notable. Chinese open-weight models have been a big story this summer on Wall Street and across Silicon Valley. The on-device Muse Glimmer model arrives less than a week after Meta debuted its first AI coding agent , known as Muse Code, to compete against the likes of OpenAI's Codex and Anthropic's Claude Code. Shares of Meta were fractionally higher in afternoon trading, at roughly $593 apiece. The stock has fully erased its late July sell-off on earnings, partially benefiting from the clearing event that was the forced unwind of the Situational Awareness hedge fund and Amazon CEO Andy Jassy's thoughtful defense of aggressive AI spending. While we still believe that Zuckerberg could've articulated a better AI strategy on the company's Q2 earnings call, we welcome these positive incremental updates on Monday. After the closing bell on Monday, mall owner Simon Property Group and infrastructure consultancy AECOM are set to report earnings. Before the opening bell Tuesday, we'll see earnings from Club name Cardinal Health , On Holding and Aramark . On the economic data side we'll get the NFIB small business optimism index and July existing home sales. (See here for a full list of the stocks in Jim Cramer's Charitable Trust.) As a subscriber to the CNBC Investing Club with Jim Cramer, you will receive a trade alert before Jim makes a trade. Jim waits 45 minutes after sending a trade alert before buying or selling a stock in his charitable trust's portfolio. If Jim has talked about a stock on CNBC TV, he waits 72 hours after issuing the trade alert before executing the trade. THE ABOVE INVESTING CLUB INFORMATION IS SUBJECT TO OUR TERMS AND CONDITIONS AND PRIVACY POLICY , TOGETHER WITH OUR DISCLAIMER . NO FIDUCIARY OBLIGATION OR DUTY EXISTS, OR IS CREATED, BY VIRTUE OF YOUR RECEIPT OF ANY INFORMATION PROVIDED IN CONNECTION WITH THE INVESTING CLUB. NO SPECIFIC OUTCOME OR PROFIT IS GUARANTEED.

Meta's 'open source' Muse Glimmer model can run on a single computer - Engadget

The download is available for free and users can run the model on their own PCs. Meta has released a new slimmed down "open source" AI model that's light enough to run on a single computer, the company announced today. Called Muse Glimmer, it's based on Meta's Spark 1.2 closed model, but is small enough to require just a single GPU for agent-oriented tasks like scheduling and file management. "We designed Muse Glimmer to balance capability against the memory and compute constraints of local hardware," the company wrote. Facebook said that it's making the "weights" that AI systems use to choose responses available to everyone on Hugging Face along with developer documentation. The download is available for free, and users can run the model on their own PCs. The company noted that optimized integrations will land on llama.cpp and other sites, "so you can go from download to working agent in minutes." The model is powerful for its size, according to Meta, with the "strong success rates" on benchmarks like DeepSearch QA, MCP-Atlas and SWE-Bench (which evaluates its ability write and debug code). It also supports reliable tool use, multi-step reasoning, failure recovery, multimodal input and scaffold compatibility for work with OpenClaw and other agent orchestrators. It was trained on data from over 100 languages, the company added. Meta's new model is not unlike some AI competitors from China like DeepSeek, which can run on local machines and offers more open licensing. "Rather than centralizing superintelligence, we should distribute it widely and give every person the ability to direct it," CEO Mark Zuckerberg said in an essay accompanying Muse Glimmer's release. Meta's Muse Spark AI is seen as weaker than rivals from OpenAI and Anthtropic. The new model, then, looks like an effort to take its AI in a more open-source direction and attract users who prefer to run their models locally. That has been a hot subject of late, with NVIDIA recently launching the Open Secure AI Alliance to improve cyber defense following the rogue attack on Hugging Face by an unreleased, closed OpenAI model.

Meta Unveils 'Open Source' Version of Its Most Powerful A.I. Model

Last month, Mark Zuckerberg, Meta's chief executive, boldly came out in support of "open source" artificial intelligence models that are publicly available to download and modify without payment or approval. The way forward, he said, was through "openness and putting the power of tech into more people's hands, not fewer." On Monday, Mr. Zuckerberg doubled down on that support. Meta released an open-source version of its most powerful A.I. model, Muse Spark. Called Muse Glimmer, it is nearly identical to Muse Spark and can generate code, text and images. Muse Spark, which debuted in July as a "closed" A.I. model that people pay to access, will remain closed. The new open-source model was accompanied by a 14-page essay from Mr. Zuckerberg laying out his vision for A.I. The technology should not be feared, he said, and the United States needs to retain its primacy over China in the A.I. race. Without naming Anthropic and OpenAI, the rival A.I. companies that have called for tight controls on the technology, he criticized their approaches, saying that concentrating A.I. power in the hands of just a few companies was hazardous. "Rather than centralizing superintelligence, we should distribute it," Mr. Zuckerberg, 42, wrote in the essay, which he titled "The Future is for Everyone." He added, "This has the potential to begin a new era of personal empowerment." Meta's new model is likely to add to a contentious Silicon Valley debate over the direction of A.I. development. Anthropic and OpenAI have argued that A.I. should be restricted because it is dangerous and needs to be safely developed in a contained fashion. But Meta, Nvidia, Microsoft, Google and others have backed open-source, saying the technology must be freely available so that people can further develop A.I. and build new businesses. The debate was turbocharged last month when Chinese start-ups released powerful open-source A.I. models that appeared to threaten the technological edge held by U.S. companies. Officials in Washington became embroiled in the debate, with OpenAI and Anthropic raising concerns about Chinese open-source A.I. models with regulators, while Nvidia and others pushed against restrictions. (The New York Times has sued OpenAI and Microsoft, claiming copyright infringement of news articles. The two companies have denied the claims.) Meta has long supported open-source A.I. and was for years one of the few American tech giants to give away its top A.I. models for free. That changed last spring, after Meta fell behind Anthropic and OpenAI in the A.I. race. It began developing a closed A.I. model instead. That model, Muse Spark, was developed under Alexandr Wang, 29, an A.I. entrepreneur who Meta hired last year to be its chief A.I. officer. The company began selling access to Muse Spark for the first time last month. Last week, Meta also released Muse Code, a stand-alone coding assistant powered by Muse Spark. Muse Spark does not perform as well as leading A.I. models from Anthropic and OpenAI on measures such as coding, reasoning and writing, according to leaderboards that track A.I. benchmarks. Meta has positioned the model as a cheaper alternative to its rivals. That strategy is similar to Chinese A.I. companies and their inexpensive open-source models. The Chinese models have become increasingly popular in the United States as companies seek to cut their A.I. bills, as well as in developing nations such as Kenya and Uganda. In his essay, Mr. Zuckerberg also expanded on Meta's vision for superintelligence, which includes giving people personal "agents," the A.I. bots that can "work 24/7 on your behalf to improve your relationships, health, career, finances, home management, hobbies, and more." His eight-year-old daughter has been using A.I. to code and create videos, he added, and he is using A.I. with her to design a robot. "Everyone will soon have invention superpowers," Mr. Zuckerberg said. China has advantages in the A.I. race, Mr. Zuckerberg added, such as the ability to build data centers and nuclear power facilities quickly and with fewer regulations. He called for a new process for government oversight of A.I. development, and defended distillation, a practice in which A.I. labs use rival models to help train their own. In the coming months, Meta plans to release a new cutting-edge A.I. model, internally called Watermelon, that will be more powerful than Muse Spark. Mr. Zuckerberg did not say if the model would be open or closed.

Meta's latest AI model wants to live on your PC

The model is available now on Hugging Face, with support for popular local AI tools coming soon. Most modern, capable AI models rely on cloud computing to give you fast and reliable responses -- your prompt and associated data has to leave your device, head to a server, get processed by the model in use, and then make its way back to you as a response. But we've also seen models taking an on-device approach, one that minimizes concerns related to cloud-based AI, but at the same time introducing their own limitations. A new model from Meta takes the latter approach, and it does so in a way that tackles some of on-device AI's main constraints head on. With cloud-based AI, one of the biggest limitations is that models cannot run without an active internet connection. Models like Meta's new Muse Glimmer that run totally on-device don't face this limitation. Then there's the privacy concern with queries going to the cloud. You're simply trusting the company behind the AI tool with your data every time you send in a request. This isn't a concern when you opt for the on-device approach. To be clear, Muse Glimmer isn't the first model to break away from the cloud server approach. Google's Gemini Nano and Gemma 4, Microsoft's Phi-4-mini, and even Meta's own Llama 3 can run locally. However, said models are extremely lightweight (at least when compared to Meta's new model) and focus on simpler tasks. Muse Glimmer, in comparison, focuses specifically on agentic AI. That's what makes its on-device existence so special. Muse Glimmer itself isn't necessarily lightweight. It is a 30-billion-parameter model. Gemini Nano 1, for comparison, has 1.8 billion parameters, while Nano 2 has 3.25 billion parameters. Nano 3 and Nano 4 go up to roughly 4 billion parameters. The tech giant says that a model like Muse Glimmer would normally require over 55GB of memory. Meta gets around this memory barrier using 4-bit quantization, bringing the model down to under 20GB. "It's small enough to run on a Mac or PC with a single consumer GPU, enabling use cases that range from local agents and function calling, to local coding, and LLM-as-a-judge evaluation," wrote the company. Muse Glimmer can write and debug code, resolve multi-turn commands from start to finish, and work through long tasks without losing track of the task or context. In case something goes wrong, the model is capable enough to "diagnose the error and retry rather than halt." Additionally, the model is trained on data from over 100 languages, and it can toggle between different reasoning strengths to hit a balance between quality and speed. Oh, and it's fast. For those unfamiliar, regular AI models generate their responses one token at a time. This can result in a delayed response when the model has to perform a complicated task. With Muse Glimmer, Meta is taking a slightly different approach. Using a workaround called speculative decoding, Muse Glimmer is able to leverage a lightweight "drafter" model based on DFlash, which is a technique used to speed up AI chatbots' response by using prediction. The drafter model's only job is to quickly predict what Muse Glimmer is likely to generate next. Essentially, the drafter is able to propose several contextual tokens at once, which Muse Glimmer can then accept or reject. This results in a faster overall response time without compromising output quality. Meta's Muse Glimmer is available now on Hugging Face. Support for Ollama, LM Studio, Unsloth, llama.cpp, ExecuTorch, and MLX is on the way.

Mark Zuckerberg's Answer to Growing AI Safety Concerns Is to Just Trust People to Do the Right Thing

Meta is stepping up to lead us all into a bright and glorious future, where democracy reigns and everyone has access to a completely secure, constantly available AI assistant to help them invent pretty much whatever they want. That's at least what Meta CEO Mark Zuckerberg has promised in a lengthy new essay, published this morning alongside the release of Meta's latest AI model -- Muse Glimmer -- and less than a week after the company was fined $567 million by a New Mexico judge who blamed it for causing "a public nuissance" within the state. The essay, running to around 6,500 words and titled "The Future is for Everyone," is a not-so-subtle dig at Meta's primary competitors, OpenAI and Anthropic. While Zuckerberg never explicitly calls out those two companies by name, he argues that the proprietary AI systems like the ones they're known for (ChatGPT and Claude, respectively) threaten to usher the world into an era in which a tiny handful of corporate giants control the rest of humanity's access to "superintelligence" -- his preferred term for future AI systems. The antidote to such a dystopian oligarchy scenario, he argues, is to make AI freely available to as many people as possible, specifically through the free proliferation of open source models. "Rather than centralizing superintelligence, we should distribute it widely and give every person the ability to direct it," Zuckerberg wrote in his essay. "This has the potential to begin a new era of personal empowerment where individuals can use this powerful new capability to reach their full potential, pursue their interests, and improve their lives and the world more than ever before." A new age for invention AI, he writes, will soon enable invention on a mass scale. Whereas major technical breakthroughs of the past were achieved only sporadically, soon they'll become commonplace: "Everyone will soon have invention superpowers." Sounds great, if human beings only invented things that benefited themselves and others, which is very rarely (if ever) the case; every technology -- from writing, to the steam engine, to radios -- has a dark side. Modern AI systems, which can be used to spread misinformation on an unprecedented scale and synthesize completely new viruses, can be especially dangerous. Zuckerberg's essay pays lip service to the litany of concerns that many people have about the spread of AI, and responds by essentially arguing that the best way to deal with any hiccups that might arise is to trust people to do the right thing. "The arc of human civilization has bent towards putting more power in people's hands to live and shape the world in the ways we believe are best," he writes. In response to fears that AI will displace huge numbers of humans in the job market, for example, Zuckerberg points to the Industrial Revolution, which "steadily freed much of humanity to focus less on subsistence and more on the pursuits we choose," he writes. "At each step, people used our newfound productivity to achieve more than was previously possible, as well as spending more time on creativity, culture, relationships, and enjoying life." The implication is that AI will achieve much the same thing: By liberating us from the drudgery of mundane tasks, we'll be free to focus on more meaningful, fulfilling pursuits. It's a common refrain from Silicon Valley, one that tends to conveniently overlook the long shadow of industrialization: growing wealth disparity, ecological catastrophe, mechanized warfare, colonization, and so on. Trust the free market Meta has long sought to position itself as a champion of open source AI, which unlike proprietary models can be freely accessed and modified by third-party developers. Muse Glimmer, the new model released today by Meta, is not open source. It's what's known in the industry as "open weight," meaning while external developers can view the parameters that were used to train the new model, they can't actually view the underlying source code or use them to build a new model from the ground-up. As the psychologist (and outspoken AI industry critic) Gary Marcus put it in an edition of his email newsletter this morning: "open weights is open source with all the good press but far fewer advantages." In his essay, Zuckerberg repeatedly frames open source as humanity's best chance of avoiding the most serious risks posed by AI, including not just mass job displacement but the rise of AI-assisted totalitarianism and civilization-scale destruction. Again, his argument boils down to the laissez-faire ethos of a free market. "A healthy balance of power is to ensure that there is no singular centralized superintelligence," he writes, "but instead as many people and businesses as possible with different superintelligent agents aligned to their goals that check and compete with each other in the ways our natural economy behaves." Open source AI has become a political flashpoint in recent months in the wake of a flurry of new AI models from Chinese AI labs, many of them open source, that rival the capabilities of the most advanced American-made models. The Trump administration and some US tech leaders have accused Chinese AI developers of large-scale distillation of American models. Zuckerberg says that any federal crackdown on distillation would be a mistake, and that all publicly available tools are fair game for the development of new models. "This is how the world works," he wrote. In a closed-door meeting with American tech executives last week, White House officials reportedly shared a draft of their new framework for testing the safety of unreleased AI models. Open source models are currently exempt.

Meta launches Muse Glimmer and presses Washington on open-weight AI

Muse Glimmer is a 30-billion-parameter model, released under Apache 2.0, that runs on a single consumer GPU, and its launch doubles as Zuckerberg's pitch for Washington to give American open-source AI a freer hand. Meta has released a new AI model, and Mark Zuckerberg has used the moment to make a political argument as much as a technical one. The model is Muse Glimmer, a compact, open-weight system, and its arrival came wrapped in a call for Washington to stop holding American open-source AI back. The model itself is built for a specific kind of user. Muse Glimmer is a 30-billion-parameter system aimed at agentic tasks, and crucially it runs on ordinary consumer hardware, a Mac or PC with a single graphics card, which puts it squarely in the growing camp of on-device AI rather than the cloud-bound giants. Its efficiency and openness are the headline features. After 4-bit compression the model needs under 20GB of memory, so it fits on a 24GB or 32GB card and runs on hardware as ordinary as a MacBook or a single RTX 5090 Also, Meta is releasing it under the permissive Apache 2.0 licence on Hugging Face, with day-one support across local-AI tools like Ollama, LM Studio and vLLM. On capability, Meta pitches it as a genuine agent rather than a chatbot, highlighting reliable tool use and function calling, multi-step reasoning over long tasks, the ability to recover from its own errors, and multimodal input, all with a dial for how much reasoning effort to spend. The benchmarks it cites are revealing about its ambitions. Meta compares Muse Glimmer with Google's Gemma and Alibaba's Qwen rather than with the frontier giants, positioning it as a leader in its size class and a bet that the future is not only enormous cloud systems but also smaller models that live on your own machine. Zuckerberg's framing, though, was where the news really sat. He argued that US policy has to change if American open-weight models are to lead over time, insisting that domestic labs are being handicapped by rules their foreign rivals simply do not face. His specific complaint was about training data. American labs, he said, must comply with many additional restrictions on the data they train on, a friction he claims gives foreign labs a structural head start in the open-weight race. The subtext is a contest Meta is worried about losing. Chinese developers now dominate open-weight AI, with models like Moonshot's Kimi K3 and Alibaba's Qwen setting the pace, and DeepSeek's V4-Flash crowned the cheapest well-known model to run, exactly the sort of efficient, downloadable system Muse Glimmer is meant to answer. Mark Zuckerberg was pointed about how not to respond. Rather than trying to ban or restrict foreign open-source models, which he dismissed as ineffective, he wants the US to level its own playing field, and he openly championed distillation, the technique of shrinking large models into smaller deployable ones. Muse Glimmer is that philosophy made concrete, distilled from Meta's larger Muse Spark to pack a bigger system's know-how into one small enough to run at home. That endorsement is notable given where the industry lines are drawn. American frontier leaders, including OpenAI, Anthropic and Google, have largely kept their most powerful models closed, so Meta's loud embrace of open weights sets it apart from its closest domestic rivals. There is a tension in Meta's own stance, though. Zuckerberg had earlier hinted that the company might be more cautious about open-sourcing its most powerful superintelligence systems, so the open banner now flies most confidently over smaller models rather than the frontier. The push also lands amid a broader argument about who leads on openness. A group of AI figures has pressed the case that open weights are central to American AI leadership, a debate in which Meta has positioned itself as the standard-bearer among the big US labs. All of this sits inside Meta's enormous AI bet. The company has been lifting its capital spending to build the infrastructure for superintelligence, and open models are the part of that strategy designed to win developers and goodwill rather than direct revenue. For now, Muse Glimmer is both a product and a lobbying prop. It gives developers something genuinely useful to run at home, and it hands Mark Zuckerberg a fresh example to wave at Washington as he argues that the way to beat China on open AI is to loosen the reins on America's own labs.

Meta to open source its most powerful AI model as it takes swipe at OpenAI, Anthropic

Meta is looking to reassure investors that it is competing with leading labs like OpenAI and Anthropic. Meta said it would open source its most powerful AI models and launch new ones designed for consumer devices, as it looks to rival leading labs like OpenAI and Anthropic. In an Instagram video posted Monday, CEO Mark Zuckerberg said that the company would open the weights for its latest AI model Muse Spark 1.2, meaning it can be downloaded and used by the public. Weights refer to the calculations and rules that determine how the AI works and behaves. Zuckerberg also said the company will release a new family of open-source models called Muse Glimmer, designed to run on a laptop. Shares of Meta were up 2.1% in premarket trade on Monday. They are down around 10% so far this year, as investors scrutinize Meta's spending plans and whether it can compete with major AI leaders. Zuckerberg is looking to reassure investors that the Meta Superintelligence Labs, formed last year, is making progress and the capital expenditure, forecast to be up to $145 billion this year, will pay off. While companies like Anthropic and OpenAI have mainly focused on closed AI models, rivals in China from Alibaba to DeepSeek and Moonshot have aggressively released open-weight models that have competed in some areas with the leading technology out of U.S. companies. In a 6,500-word essay on AI, published Monday, Zuckerberg positioned Meta's move as a challenge to Chinese open-source technology and urged Washington to support American efforts. "Foreign labs currently hold several advantages here since American labs have to comply with many additional restrictions on training data. US policy must reduce this additional friction if we want American open source models to lead over time," Zuckerberg wrote. "I do not believe restricting access to foreign open source models is an effective solution. Our goal should be for American open source models to be the best globally. This requires removing the hurdles that make it harder for American open source models to compete." Zuckerberg said that the U.S. will need to "rethink" its policies in several areas, including distillation and data use in training, if it is to lead in open source AI. Broadly, distillation refers to the use of output from an advanced AI model to train another model. Distillation has attracted some controversy, with some U.S. lawmakers viewing the practice as effectively theft of intellectual property. In touting his open-source push, Zuckerberg warned against the concentration of power in AI around a few companies, in what appeared to be a thinly veiled swipe at OpenAI and Anthropic. "Still, it is surprising that the discourse from many developing AI is so filled with doom. I do not understand why anyone who believes that AI will eliminate most jobs and much of humanity's relevance would rush to build that future," Zuckerberg wrote. "The notion that AI is so dangerous that the only safe path is an extreme concentration of power seems inherently problematic." Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman have both warned of AI's impact on jobs. Altman recently softened some of his comments on jobs. Choose CNBC as your preferred source on Google and never miss a moment from the most trusted name in business news.

Mark Zuckerberg Posts Deranged 6,500-Word Essay About Giving Everyone AI Superintelligence

"The future is for everyone," Zuckerberg says, describing future that is primarily good for Meta. Mark Zuckerberg, whose superyacht apparently spent the weekend ignoring or missing the distress signal from a boat that ran out of fuel near Alaska, has posted a deranged, 6,500 word essay detailing his vision for AI superintelligence, a future that is "for everyone" but which sounds less social than ever. Zuckerberg posts these types of essays every so often for purposes that serve his own company, and this one, called "The Future Is For Everyone," is designed to defend against general backlash to AI but also to Meta's own practices. Zuckerberg lays out the potential use case for Meta glasses (whose huge marketing campaign cannot get people to stop calling them "pervert glasses"), AI agents, open weights AI development, and why data centers are not bad for communities, actually. Like most Silicon Valley "utopian" essays, to believe that any of this is going to go how Zuckerberg suggests it will requires one to have been recently concussed or to willfully ignore how this technology is being used today and believe that thousands of years of human nature will suddenly shift. For example, Zuckerberg writes "Everyone will have an exceptionally capable personal agent that understands you, your goals, and everything you care about. Your agent will work 24/7 on your behalf to improve your relationships, health, career, finances, home management, hobbies, and more. It will free up time for the things you enjoy, and help you accomplish more than you could otherwise. It will have strong privacy and security options so you can trust it to handle all of your personal content knowing that no one else can access your information, similar to how encryption works on WhatsApp. You'll be able to interact with your agent through any device, including your glasses to keep you present in the moment with the people you care about." Zuckerberg does not grapple with, or even gesture at, the idea that some people may not want to have an AI agent working on their "hobbies." He does not consider that, even if everyone were to have an AI agent, perhaps not everyone would use these AI agents for good. In the few months that AI agents have become popular among the early adopter set, we have seen "benevolent" AI agents endlessly spam humans and the internet with drivel. And those are just the kind-of-annoying ones. We have seen AI agents hack companies, and over the weekend an Australian man went viral because his AI agent that he asked to sign him up for gym classes did so by hacking the gym's reservation system and canceling other people's reservations. Like many AI weirdos, Zuckerberg explains how AI has already changed his life by allowing him to automate many of the joyful tasks of parenting by outsourcing them to an agent: "My agent flags interesting information and helps me prototype ideas. It helps keep me healthy by monitoring my sleep and then watching as I train and giving feedback. My daughter loves to bake so my agent plans personalized recipes for us to make together each weekend, orders the ingredients, and then offers suggestions as we're baking," he writes. "My 8 year old daughter can already code her ideas and produce videos in an evening that would have either taken me months or been impossible previously. Now we're designing a robot together." Zuckerberg's essay goes on and on and on like this. He imagines a future where everyone can do everything and wants to do everything. In Zuckerberg's future, everyone will have a business run by their AI agent. Everyone will be inventing things and doing basic research on the nature of the universe and physical elements, for some reason. Zuckerberg writes that Meta employees, with the help of AI, are "generating novel crystal structures that are ideal for augmented reality glasses," then writes "everyone will soon have invention superpowers," and that "everyone [...] will be able to contribute to scientific progress." In this future that, again, is for "everyone," Zuckerberg explains that AI tools will be free, but that, actually, using it will be a tiered system that is exactly the same as it is now: "For everyone to be part of the future, everyone must have the ability to use superintelligence to improve their lives and shape the world. We will offer free versions that will be accessible to billions of people. For those who want to pay to use more compute, there will be a dynamic auction mechanism that will guarantee that everyone gets the lowest price possible for the intelligence and compute they're using while also ensuring the capacity is used for whatever people collectively find most valuable." Zuckerberg's essay is full of platitudes and sentences that mean nothing, "thought experiments" that are not developed or explored in any way, discussions of "freedom," etc. Here are some sentences: * "Humanity is not a monoculture. People's diverse values represent different tradeoffs they would make on important issues. There is no technological solution that can align with everyone's opposing interests and values at once." * "As a thought experiment, imagine only one person had a superintelligent lawyer. They would have an unfair advantage in court -- even if they were wrong on the merits. That would lead to a worse society. But now imagine everyone has a superintelligent lawyer. In this case, justice would be carried out much more fairly and efficiently than it is today when there is often an imbalance in skills and resources in litigation." * "People have an infinite demand for new experiences and have always found new problems to tackle." * "While the number of questions a person can ask in a day is limited, the number of valuable things superintelligence can invent to help achieve your goals is unlimited." * "If people can use AI to invent incredibly valuable new things, then it will make more sense to allocate it towards that rather than automating existing jobs. The more superintelligence serves as a tool of invention, the more likely that individual capability outpaces automation and the future is better for people." * "In a free society, people will have tools that can be used for good or harm, but law enforcement and military have more weapons and intelligence-gathering." There is an entire section on data centers. In Zuckerberg's future they are powered by power infrastructure Meta owns (which is not the case currently) and create lots of high-paying, long-lasting jobs (not the case currently). There is an entire section on "preventing government tyranny" by giving everyone superintelligence: "To maintain freedom, we must ensure that superintelligence primarily empowers individuals. The ideal in liberal democracy is that people naturally hold all rights and only agree to restrict some freedoms to protect the common good. Similarly, individuals should have access to personal superintelligence and should only be subject to restrictions when truly required." Zuckerberg does not address the backlash to his company, his data centers, his social media platforms, or his surveillance glasses. He does not discuss the slopification of the internet, gestures at job loss only through the lens that superintelligence will somehow fix it once AI agents start businesses for everyone or "invent incredibly valuable new things," and describes a future in which bad actors essentially do not exist or are easily dispatched with. It's the future Zuckerberg wants. It's not the future "everyone" wants.

Meta launches Muse Glimmer AI model for laptops

Meta thinks the next front in the AI race may be sitting on your desk. On Aug. 10, Meta released Muse Glimmer, a 30-billion-parameter AI model designed to run locally on a Mac or PC with a single consumer GPU. Its weights are available to download from Hugging Face under a permissive Apache 2.0 license, allowing developers to use, modify, and build products with the model for free. That means Muse Glimmer can operate without sending every request to a distant data center. It can also work without an internet connection, provided the computer running it has enough memory and processing power. For anyone ready to open an app and start chatting, however, there is a catch: Muse Glimmer is a developer release. Meta is supplying the model that can power an AI agent, rather than a consumer product that immediately manages your inbox or reorganizes your desktop. What can Muse Glimmer actually do? Meta built Muse Glimmer for "agentic" work, meaning it can take a goal, divide it into steps, and use connected tools to complete parts of the task. The company imagines local agents that could manage schedules, draft messages, organize files, and write or debug code. The model can process images alongside text, perform multi-step tasks, and attempt recovery when a tool fails. Meta says it was trained on data from more than 100 languages. Fitting those abilities onto one computer required some downsizing. At full precision, Muse Glimmer would need more than 55 GB of memory. Meta compressed the model to below 20 GB, leaving additional room for its working memory and image-processing system. Even so, "runs on your laptop" is doing some work here. Meta designed the compressed version to fit within a 24 GB or 32 GB memory envelope and tested it on hardware, including Apple's M4 Max and M5 Max chips and Nvidia's RTX 5090. Your average work laptop may sit this one out. For people with the right hardware, local processing offers some clear advantages. The model can work offline, developers can avoid paying for every request to a cloud service, and personal information does not necessarily have to leave the device. The level of privacy will still depend on how each application is built and which outside services it can access. In Meta's own benchmark tests, Muse Glimmer performed competitively against similarly sized models from Google and Alibaba, although those results have not yet been independently verified. Zuckerberg takes aim at closed AI Meta describes Muse Glimmer as open source, although "open weight" is the more precise term. The calculations the model learned during training are public, but the release does not expose all the data and infrastructure used to create it. Still, the Apache 2.0 license gives developers considerably more freedom than they receive from closed models available only through a company's app or paid API. Muse Glimmer is also the smaller opening act. Mark Zuckerberg said Meta will soon release the weights for Muse Spark 1.2, its more powerful foundation model. Muse Glimmer was distilled from Spark, meaning it was trained partly by learning from the larger model's outputs. Zuckerberg expanded his case for open AI in an essay titled "The Future Is for Everyone." He argued that the United States should loosen restrictions around training data and distillation so American companies can better compete with open models from Chinese developers, including DeepSeek, Alibaba, and Moonshot AI. He also took a thinly veiled swipe at companies that warn about AI's dangers while keeping their most capable models under tight control. "The notion that AI is so dangerous that the only safe path is an extreme concentration of power seems inherently problematic," Zuckerberg wrote. He did not name OpenAI or Anthropic, but the contrast was clear. Both companies keep their leading model weights private, while their executives have repeatedly warned about the economic and societal risks of more powerful AI. Zuckerberg instead argued that advanced AI should be widely distributed, with each person eventually controlling an agent that understands their goals and personal context. Releasing Muse Glimmer gives developers another option outside the paid APIs of Meta's rivals. It also gives Meta a way to influence the AI market even when people are not using one of its apps. Whether ordinary users are ready to give a local agent access to their schedules, messages, and files remains an open question. The more immediate hurdle may be finding a laptop muscular enough to run it.

Meta returns to open source with Muse Glimmer, an Apache 2.0 licensed 30B parameter LLM available now

Meta today released Muse Glimmer, a 30-billion-parameter open-weight model designed to run autonomous AI agents directly on consumer hardware -- pushing agentic workloads that normally depend on cloud infrastructure onto high-end Macs and PCs. Just as notable as what the model does is how it's licensed. Glimmer arrives under the permissive, industry-standard Apache 2.0 open source license -- the company's first fully open release since it succeeded its open-weight Llama family in April with the proprietary Muse Spark. In fact, Muse Glimmer launches today with a more permissive license than Llama ever carried. Llama's bespoke community license drew years of criticism for restrictions like its 700-million-monthly-user cutoff; Apache 2.0 has no such strings, permitting unrestricted commercial use, modification and redistribution. The weights are available on Hugging Face now. Wang said support is rolling out this week through Ollama, LM Studio, vLLM, SGLang, Together AI, Fireworks AI and OpenRouter, with optimized llama.cpp, MLX and ExecuTorch integrations landing in the coming days; Meta's blog post also names Unsloth as a local-runtime partner and points to PyTorch's TorchTitan for fine-tuning. The company says it is working with AMD, Arm, Dell, Intel and Nvidia to optimize performance across devices, and has published developer documentation covering custom agent scaffolds. "Today we're also opening the weights for Muse Glimmer, a great 30B parameter dense model that can run locally," Meta co-founder and CEO Mark Zuckerberg wrote in a post on X (under his longtime handle @finkd). "Soon we'll also release the weights for Muse Spark 1.2, our latest foundation model. Meta is a strong supporter of open source and I'm proud of these releases." That promised Muse Spark 1.2 release would be an even bigger shift: it's the frontier model behind Muse Code, the terminal coding agent Meta shipped just five days ago, and until today the entire Muse family was proprietary. Zuckerberg had teased at that launch that he'd "have more to share soon" on open source. Now we know what he meant. For developers and enterprises, the practical stakes of local inference go beyond where computation happens. An agent working with files, screenshots, development environments and other sensitive context can execute those workflows without continuously sending that information to a remote inference service. Local deployment also removes network availability and per-token API charges from the inference loop -- although organizations still bear hardware, electricity, deployment and management costs. A 30B model built around the agent loop Rather than positioning Glimmer primarily as a general chatbot, Meta trained it around the sequence of operations an autonomous agent performs: formulate a plan, call tools, interpret the results, continue working, and recover when something goes wrong. "Just like much larger models, muse glimmer can operate as a fully capable agent via planning, tool calls, checking its own results, and failure recovery," Alexandr Wang, Meta's chief AI officer, wrote in a thread on X announcing the release, adding that the model "can run on 24GB of VRAM without losing agentic reliability." According to the model card on Hugging Face, Glimmer is a dense causal transformer with approximately 29.6 billion total parameters across 52 layers, including a dedicated ~1.8B-parameter ViT-G/14 perception encoder. It accepts interleaved text and images, produces text, supports more than 100 languages and has a stated context length of 131,072 tokens or more, with a knowledge cutoff of January 4, 2026. That combination is intended to let an agent interpret screenshots, charts and documents while simultaneously reasoning about text and invoking external tools. Glimmer offers low, medium, high and xhigh reasoning settings -- set via the system prompt -- so applications can dial reasoning effort up or down per task, and Meta says it works across agentic scaffolds including OpenClaw and Hermes Agent. The model is a distillation of Meta's larger flagship: per the company's technical blog post, Glimmer was pre-trained on Muse Spark's outputs using logit distillation, mid-trained on longer-context, agent-heavy data with richer reasoning traces, then post-trained with supervised fine-tuning, on-policy distillation and reinforcement learning across general, reasoning, coding and agentic domains. Meta demonstrated the result with a local Home Assistant workflow: in a demo video, Glimmer autonomously discovers a Home Assistant instance on the network via tool calls, queries device APIs, writes a responsive HTML/CSS/JavaScript dashboard from scratch and deploys a local server to verify its own work. That's closer to the operational reality of enterprise agent deployments than a standalone question-answering benchmark -- the model has to maintain a plan while interacting with external systems, then inspect whether its actions produced the expected result. Compressing an agent into 24GB The hardware story is central to the release. At full precision, Meta says the 30B model requires more than 55GB of memory -- beyond any single consumer GPU. The company therefore developed approximately 4-bit quantized versions that shrink the language-model weights to under 20GB, leaving headroom for the pieces an operational agent also needs in memory: the KV cache, the perception encoder and a companion speculative-decoding model, all fitting within a 24GB or 32GB envelope. In practical terms, that means the quantized builds run on consumer machines -- though the upper end of them. The 24GB-targeted K-Quant-17GB configuration fits on a single high-end consumer graphics card, such as Nvidia's RTX 3090 or RTX 4090 (both with 24GB of VRAM), while the 32GB-targeted K-Quant-Dynamic version lines up with the newer RTX 5090's 32GB. On the Mac side, Apple Silicon's unified memory plays the role of VRAM, so a MacBook Pro or Mac Studio with 32GB or more of memory can hold the full stack -- Meta ran its own speed tests on M4 Max and M5 Max MacBook Pros. A typical 8GB or 16GB laptop, however, remains out of reach, and the full-precision BF16 release -- which Meta pegs at 64GB -- stays in the territory of data-center GPUs and top-spec Mac Studio configurations. Meta reports average accuracy degradation of just 0.2% across 15 benchmarks for its K-Quant-Dynamic version targeting 32GB hardware, and 1% for the K-Quant-17GB configuration targeting 24GB hardware. Those figures are Meta's own measurements, not independent evaluations. Meta is also using DFlash speculative decoding to attack the other big problem with local agents: latency. Instead of generating every token sequentially, a smaller DFlash "drafter" model proposes blocks of 16 tokens that the primary model verifies in parallel, producing identical output faster. Meta reports this raises average generation speed on an Nvidia RTX 5090 from 74.9 tokens per second to 233.4 -- a 3.1x increase. An Apple M5 Max rises from 26.6 to 50.2 tokens per second (1.8x), and an M4 Max from 23.7 to 37.8 (1.5x). The tests used batch size one and greedy decoding, with Apple systems measured via ExecuTorch and the RTX 5090 via llama.cpp. For agent applications, those multipliers matter more than they would for chat: a single user request can trigger many model turns, tool calls and verification steps, and latency accumulated at every stage can quickly make an otherwise capable agent impractical. Glimmer enters an increasingly competitive local-model market Meta is not entering an empty field. Developers already have capable open-weight models in this size class, most prominently Google's Gemma 4 family and Alibaba's Qwen3.6-27B -- both of which position themselves around reasoning, multimodal understanding and agentic workloads. Meta's own benchmark table compares directly against both. Glimmer leads that three-way comparison on several agentic tests, including MCP Atlas at 75.5, DeepSearch QA at 74.6, τ³-Banking at 23.5, WildClawBench at 47.6 and GAIA2 at 43.3. It scores 51.2 on SWE-Bench Pro, versus 36.9 for Gemma4-31B and 50.2 for Qwen3.6-27B in Meta's evaluation. But Glimmer does not sweep the field. Qwen leads Meta's own comparison on OSWorld-Verified (75.6 vs. Glimmer's 65.9), TerminalBench 2.1 (60.7 vs. 51.7), SkillsBench, GDPval-AA (1141 vs. 953) and most of the multimodal benchmarks. On SWE-Bench Verified, Glimmer's 76.0 lands just below Qwen's 77.2. Gemma leads on GPQA Diamond and Humanity's Last Exam. Read honestly, the numbers make Glimmer more interesting as a specialized local-agent model than as evidence of a universal performance lead. For enterprise developers, the practical question is whether its combination of agent reliability, quantization quality, tool compatibility and decoding speed translates from benchmarks into sustained real-world workflows. Meta Glimmer adds to a still-small roster of genuinely open, frontier-class models from U.S. companies. For the last two years, Chinese companies have set the pace in open source AI, with DeepSeek, Alibaba's Qwen team, Moonshot AI's Kimi, Zhipu's GLM and MiniMax shipping frontier-class open models under MIT and Apache 2.0 licenses on a cadence Western labs haven't matched. The usage data reflects it: by May 2026, Chinese open-weight models accounted for roughly 61% of all tokens consumed on OpenRouter, with four of the five most-used models coming from Chinese labs -- while Meta's Llama, the prior open-weight leader, fell off the rankings entirely. The U.S. counterexamples remain countable on one hand: OpenAI's gpt-oss-120b and gpt-oss-20b, released under Apache 2.0 in August 2025 as the company's first open weights since GPT-2; Google's Gemma family, which is open-weight but ships under Google's own more restrictive custom license rather than an OSI-approved one; and Thinking Machines' Inkling. Glimmer invites the most direct comparison to gpt-oss: both are Apache 2.0, both offer adjustable reasoning effort, and both target self-hosted deployment. But the gpt-oss models are text-only, sparse mixture-of-experts designs built primarily for reasoning and tool use -- gpt-oss-20b fits in about 16GB of memory while gpt-oss-120b targets a single 80GB data center GPU. Glimmer stakes out different ground: a dense model with native vision input, trained end-to-end around the agent loop, shipping with its own quantized variants and speculative-decoding drafter tuned for 24GB consumer machines. And if Zuckerberg follows through on opening Muse Spark 1.2's weights, Meta would put an actual U.S. flagship frontier model into open circulation -- something no American lab has done at that tier. Safety remains part of the deployment architecture Giving a local model access to tools creates a different security problem from deploying a local chatbot -- and Meta's own safety numbers show Glimmer is not uniformly stronger than its peers. On CI Memories, a privacy benchmark where lower violation rates are better, Glimmer records 26.4 against Gemma's 12.1 and Qwen's 53.4. On Siren AgentDojo, a prompt-injection test, Glimmer shows a 28.4% attack-success rate versus 25.6% for Gemma and 40.3% for Qwen -- while posting the highest utility score of the three at 94.2. Meta says it evaluated Glimmer under its Advanced AI Scaling Framework and determined the model does not meet the framework's definition of "Frontier AI" because it is generally less capable than Muse Spark. Its Preparedness Team assessed Glimmer at Moderate or lower risk across chemical/biological, cyber and loss-of-control categories -- the latter two inferred from the fact that Glimmer is broadly weaker than Muse Spark 1.0, which received the same designations. The company nevertheless recommends deploying Glimmer as part of a broader system with guardrails, including human-in-the-loop confirmation for irreversible actions. That caveat matters especially for local agents: keeping data on-device reduces exposure to cloud infrastructure, but local execution does not by itself solve prompt injection, excessive permissions or an agent taking an unintended action. Apache 2.0 weights and a fast-growing runtime ecosystem Meta is releasing full-precision BF16 weights, both 4-bit quantized variants, the DFlash drafter and the perception encoder -- all under Apache 2.0. There is no Meta API price attached to the downloadable model, leaving total cost dependent on local hardware or whatever third-party hosting developers choose. One nuance worth noting for procurement teams: as with most "open source" model releases, it is the weights that are open -- Meta has not released the training data or training code. The broader implication is that Meta is treating the developer workstation as a credible deployment target for autonomous agents, rather than merely a place to experiment with smaller language models. Glimmer's 30B size and 24GB target put that proposition within reach of high-end consumer hardware, while the Apache 2.0 license gives developers -- and their legal departments -- unusual freedom to modify and deploy it. The next test is whether its benchmark advantages survive the messier conditions of real software repositories, enterprise tools and long-running agent sessions. If they do, the most consequential part of Glimmer may not be another set of benchmark scores -- it may be that a class of agent previously expected to live behind a cloud API can increasingly live, and work, on the machine sitting under a developer's desk.

Zuckerberg: AI's biggest risk is one entity with too much control

Why it matters: Zuckerberg's flag in the ground, with a utopian vision of a highly personal AI that fuels individual achievement, comes as policymakers debate how much control they should have over increasingly powerful models. What they're saying: "One of the ways to achieve both individual empowerment and checks and balances is to make sure you put the technology in everybody's hands," Zuckerberg told Axios' Mike Allen in an interview ahead of the manifesto's release. * "It's definitely a different view than what you're hearing from a lot of other people in the tech industry right now." * He wrote that he wants to ensure the U.S. and democratic countries lead, and ultimately make sure "humanity maintains control over superintelligence so it serves rather than endangers us." The big picture: Even as the world confronts AI that is breaking out of sandboxes, comparing notes on message boards and creating new viruses, Zuckerberg portrays it instead as an analog to other disruptive technical innovations. * "Humanity has witnessed many transformative advances," he writes in the manifesto. "Each time there is fear that people will be left behind. But each time humanity has come out with more people sharing greater prosperity, health, and freedom." * "We believe this will be true with AI as well, and the abundance of the future can be shared by everyone," he says. "We also believe that the values that got us to this point -- like liberty, open inquiry, free enterprise, and equal opportunity -- are also the right values to build a positive future." Threat level: Zuckerberg contrasts this vision with one in which only some people, some businesses or some countries have access to the most powerful AI, suggesting that would represent a far more dangerous outcome. * "I'm personally more worried about centralization than I am about any of the specific risks others are talking about," he told Axios. * The manifesto argues that too much intervention by the U.S. could hand an advantage to other regimes -- presumably China -- especially when technology leads can be measured in weeks. * "Any policy that slows American model releases -- even by a month -- could add significant risk to American leadership while letting foreign models race ahead," he wrote. Zoom out: Zuckerberg's plan for Meta is that "everyone will have free or affordable access to these tools," with the paid versions using a "dynamic auction mechanism" to get the compute they need at the lowest possible price. "This will ensure the benefits of superintelligence are distributed widely." * He also said Meta's future AI agents will have privacy protections that are far greater than those currently offered by Meta AI. * "We believe that people should have a fully private mode for personal agents where even Meta or any other service provider cannot see or grant access to your information." * Meta AI's current policies allow the company to use nearly every interaction one has with the service to train its AI systems, though it recently added incognito mode. Between the lines: Zuckerberg also said Meta will resume releasing some open-source AI models, now that Meta Superintelligence Labs is up and running, as Axios reported in April. * Meta says it's opening the weights of a new model, Muse Glimmer, billed as "one of the highest-performing models of its size," and will open the weights for a version of the foundational model Muse Spark 1.2 in coming weeks. * Zuckerberg added that Meta's independent board will have the power to approve model-release safety criteria and review whether releases meet those standards. * And he promised a "community compact" around Meta's data center development, including local jobs, investments in schools and public services, and commitments on energy prices and water use. Local spending: Hoping to tamp down data-center backlash, Zuckerberg also announced Meta is launching the Future is for Everyone Fund "to support each community we work in directly." The fund, seeded with $1 billion for U.S. communities where Meta owns and operates data centers, will benefit teachers, first responders, and energy and water infrastructure. * A Meta training program for workers in skilled trades, with guaranteed jobs at the company's data centers -- America's Workforce Academy, which Axios reported in June -- graduated its first class last week. The bottom line: The manifesto represents Zuckerberg's clearest effort yet to weigh in on AI's future, both as a matter of policy and technology.

Meta's new AI model runs entirely offline, but your GPU needs to keep up

Meta drops a free 30 billion parameter AI model that lives entirely on your desktop. Meta has a new AI model out, and for once, the biggest headline isn't about capability; it's about freedom. Muse Glimmer, sitting around 30 billion parameters, ships with an Apache 2.0 license, meaning the weights on Hugging Face are yours to download, modify, and build on top of, no permission needed. You can simply run it on a graphics card on your local machine, no server farm or online connectivity required. Meta's Superintelligence Lab took its larger Muse Spark and essentially had it teach a smaller, leaner version to think as it does. Muse Glimmer accepts both text and image inputs, though it only answers in text. It supports over 100 languages, remembers conversations stretching past 131,000 tokens, and its knowledge stops at January 4, 2026. How much RAM do you need? At full precision, Muse Glimmer would eat up 64GB of video memory, way more than most people have lying around, especially in this AI-inflated RAM pricing age. Meta gets around this with quantization, a method that compresses the numbers the model uses so it takes up far less space. That shrinks the language portion to under 20GB. Two versions are available. K-Quant-Dynamic requires 32GB and barely loses any accuracy (0.2%), while K-Quant-17GB squeezes into 24GB with a slightly bigger, 1% accuracy hit. In real terms, that means you need at least an RTX 5090, RTX 4090, RTX 3090, or a Mac with an Apple Silicon Max chip. Does it actually feel fast? Meta paired the model with an accelerator called DFlash that predicts multiple tokens at once instead of one at a time. On an RTX 5090, that pushes speeds from 74.9 tokens per second to 233.4, over three times faster. Compared to Google's Gemma4 and Alibaba's Qwen3.6, Muse Glimmer leads in planning and multi-step tasks but falls behind when it comes to actually operating a desktop. You can grab it now through Hugging Face or LM Studio if you have the compatible hardware.

Mark Zuckerberg makes his case for American open-source AI over Chinese rivals | Fortune

In the essay, Zuckerberg argues that powerful AI should not be controlled by a handful of companies, a seeming jab at competitors like OpenAI and Anthropic. Zuckerberg positions Meta's commitment to open-source as critical to challenging Chinese open-source models from DeepSeek and Moonshot which have been getting uncomfortably close to the American frontier. He also urged Washington to support American efforts. As part of this new commitment, Meta says it plans to open the weights for Muse Spark 1.2 -- meaning it will publicly release the underlying trained parameters of the model, letting anyone download, inspect, and modify it rather than access it only through Meta's own products or API -- its most advanced AI model on the market. "I do not believe restricting access to foreign open source models is an effective solution. Our goal should be for American open source models to be the best globally. This requires removing the hurdles that make it harder for American open source models to compete," Zuckerberg wrote in the essay, part of a coordinated media blitz by Meta that included a Zuckerberg-penned opinion piece on the same themes in the Wall Street Journal last month and interviews with select media outlets. The Muse Glimmer release lands in an American market increasingly dominated by closed-sourced AI companies, those that sell controlled access to their best models. OpenAI, Anthropic, and Google generally keep their most advanced "frontier" models locked behind closed doors, retaining tight control of the weights and infrastructure. Meta initially sought to differentiate itself by embracing open source, launching the open-weight Llama family of models in February 2023. The Llama family, which expanded through several versions, helped make high-performing models available for outside developers to download and modify, and gave Meta influence over the wider AI ecosystem even as rivals captured more revenue by selling access to their proprietary models. Over the past year however, Meta's commitment to open-weight has blurred, as the company mixed open releases with proprietary ones. Meta has spent heavily on its new Superintelligence Labs, including bringing in former Scale AI CEO Alexander Wang. The company has been trying to position the Muse family as the foundation for a more capable Meta AI assistant across Facebook, Instagram, WhatsApp and its Ray-Ban Meta AI glasses, while also being competitive with models from Anthropic and OpenAI. The shift and refocus on open-source may also reflect an attempt by Meta to try to regain momentum after recent models have fallen short of systems from OpenAI and Anthropic. Meta is also looking to reassure investors that it is still competitive with frontier labs. Meta said it trained its new Glimmer series using distillation, a process in which a smaller model learns from Muse Spark, its larger proprietary system. In his essay, Zuckerberg wrote that American AI developers face disadvantages against Chinese rivals, including in training data, and argued that policy should reduce those barriers. With the strategy shift, Meta will likely be hoping its commitment to open-source will set it apart from rivals that keep their most capable systems behind closed doors and give it a political argument for user and government support.

Mark Zuckerberg AI essay pushes back on doomers, open source

The Meta CEO argues for distributing AI broadly, announces a $1 billion community fund, and calls for earlier government access to models Meta $META CEO Mark Zuckerberg published a 6,500-word essay on Monday pushing back on what he described as a doom-laden discourse inside the AI industry, arguing that spreading powerful AI tools as broadly as possible -- rather than concentrating them -- is the path to both economic growth and safety. "It is surprising that the discourse from many developing AI is so filled with doom," Zuckerberg wrote. "The notion that AI is so dangerous that the only safe path is an extreme concentration of power seems inherently problematic." The essay, which also appeared in shorter form as a Wall Street Journal opinion column, advances a philosophy built around three pillars: individual empowerment, invention as AI's primary purpose, and a balance of power as the foundation for safety. Zuckerberg argued that superintelligence should be distributed to individuals rather than controlled by companies, governments, or AI systems themselves, and that this distribution is what will keep humanity safe. Among the concrete announcements in the essay, Meta said it will launch a new open-weight model called Muse Glimmer and, in the coming weeks, release an open-weight version of its most advanced model, Muse Spark 1.2, according to the Wall Street Journal. Meta had paused open-weight model releases after reorganizing its AI operations last year, according to the Journal. Meta also unveiled a $1 billion "Future Is For Everyone Fund" aimed at channeling resources to the towns and regions that host its data centers. Zuckerberg pointed to Richland Parish, Louisiana, as a model, noting that local teachers had collected $50,000 bonuses tied to economic activity flowing from the company's Hyperion data-center campus nearby. The announcement comes at a fraught moment for the industry; New York moved just weeks ago to halt new data-center construction in the state for as long as a year. On government oversight, Zuckerberg proposed that AI labs share intermediate model training checkpoints with the federal government rather than waiting for a completed model. He took issue with the Trump administration's executive order that sets up a voluntary 30-day review window for new models, arguing the timeline amounts to 'quite a meaningful amount of time' at the current speed of AI development -- an argument he first raised last month in a shorter Wall Street Journal column. Zuckerberg also said Meta would give its board of directors authority to approve safety criteria for model releases. "I do not think it is in my, Meta's, or the world's best interests for me or anyone else to be a sole decision maker on how superintelligence is deployed," he wrote, calling on other AI companies to adopt similar governance structures. The essay drew an implicit contrast with Anthropic CEO Dario Amodei, who has warned about potential job disruption from AI, though Zuckerberg did not name him, according to MarketWatch. Zuckerberg argued that wide distribution of AI is more likely to generate jobs than eliminate them, predicting more companies with smaller headcounts rather than mass unemployment.

You Can (Maybe) Run Meta's Latest AI Model Locally on Your Computer

Your computer should have anywhere from 24GB to 32GB of dedicated RAM in order to run Muse Glimmer locally. On Monday, Meta announced its latest AI model: Muse Glimmer. The company isn't necessarily advertising this model as its biggest and best, however. While Meta has plenty to say about Muse Glimmer's performance potential, there are two key selling points for the model in the company's eyes: The fact that it's open-weight, and that it's designed to run locally on "consumer" Macs and PCs -- though perhaps not on your computer. What is Muse Glimmer? According ot Meta, Muse Glimmer is a 30-billion-parameter AI model optimized for "always-on local agent workflows." What that means is Glimmer is designed to run autonomously directly on your machine, rather than outsourcing that processing to the cloud. Local AI is much more secure than cloud-based, as your data stays within your system. When you tap into cloud servers, you're at the mercy of whichever company owns that network, potentially putting your data in jeopardy. While there are many use cases for a model like Muse Glimmer, Meta is emphasizing its use for agentic AI. AI agents can perform tasks on your behalf and are growing in popularity, as users and developers can ask their agents to do any number of functions. Running agents locally is also quite popular, as evidenced by OpenClaw's infamous year, but the most demanding models also require powerful hardware. If you're working with a laptop, you might not have the energy necessary to run AI agents. That's part of Meta's pitch here. The company says Muse Glimmer was trained to balance performance with limited hardware. Meta says it puts Muse Glimmer through three sets of training to achieve this: a "Pre-Training," which trained Muse Glimmer on Muse Spark's outputs; "Mid-Training," which trained the model on more "agent-heavy" data; and "Post-Training," which fine-tuned the model across a number of parameters. Meta says Muse Glimmer can handle end-to-end agentic task completion, multi-step reasoning, recovery when a tool call fails, and multi-model input and reasoning, among other functions. It is also trained on data from over 100 languages. The company says the model performs "strongly for its size class" against many standard LLM benchmarks. While Muse Glimmer isn't as demanding as other models, it likely won't run well on the average laptop. Meta says that a 30-billion parameter model would typically require more than 55GB of memory, but through "quantization techniques," Muse Glimmer can use less than 20GB. When you take its "working memory" into consideration, Muse Glimmer will likely use 24GB or 32GB of memory. If you have a "Pro" MacBook or PC, that might suit you well. But if you're like me, your laptop has 16GB or less of RAM -- not quite what Muse Glimmer requires. Still, it's a step forward for local AI use. (For reference, Meta says it tested Muse Glimmer on MacBook M4 Max, M5 Max, and with an RTX-5090 GPU.) Meta says Muse Glimmer is "open," which means Meta released the model's training weights. As such, not only can anyone use the model, but they can tune it as well. If you have specific needs for your AI model, you can tinker with Muse Glimmer to better serve your usage. How Muse Glimmer compares to other AI models There are a lot of AI models on the market these days, but Meta only has direct comparisons with two other models: Google's Gemma4-31b and Alibaba's Qwen3.6-27B. According to Meta's benchmarks, Muse Glimmer beats both models in the following 12 tests: MCP Atlas, DeepSearch QA, 𝜏³-banking, WildClawBench, GAIA2, SWE-Bench Pro, Sci Code, Charxiv Reasoning, IFBench, AIME 2026, AA-LCR, and Beam 128K. Gemma4-31b still has the crown in four of the benchmarks, while Qwen3.6-27B leads in eight. Meta didn't provide similar comparisons with models from industry leaders, like OpenAI, DeepSeek, Z.ai, or Moonshot AI. However, Artificial Analysis has a ranking of all open-weight models on the market. As of this article, Muse Glimmer (high) is actually in 18th place. Moonshot AI's Kimi K3 (max) is in the lead, with Z.ai's GLM-5.2 (max) in second, DeepSeek's V4 Flash (max) in third, and Kimi K3 in (low) in fourth. For reference, Gemma4-31b is in 32nd, while Qwen3.6-27B is actually in 17th. There are many variables here, but this leaderboard gives us a quick glance at how Meta's latest model ranks amongst the competition. How to try Muse Glimmer You can try Muse Glimmer right now by downloading the weights from Hugging Face. Meta says the model will be available in the coming days from apps like Ollama, LM Studio, and Unsloth.

5 takeaways from Zuckerberg's essay on his vision for superintelligence

Mary Cunningham is a reporter for CBS MoneyWatch. She previously worked at "60 Minutes," CBSNews.com and CBS News 24/7 as part of the CBS News Associate Program. Mark Zuckerberg said in an essay published on Meta's website on Monday that artificial intelligence tools should be shared with "as many people and businesses as possible." In a 6,500-word post outlining the Meta founder's vision of AI, Zuckerberg also expressed his belief that "superintelligence" -- a theoretical form of AI with superior cognitive abilities to those of humans -- will advance national security, lead to greater economic prosperity, and promote scientific and creative achievement. "Humanity has witnessed many transformative advances. Each time there is fear that people will be left behind," Zuckerberg wrote. "But each time humanity has come out with more people sharing greater prosperity, health and freedom. We believe this will be true with AI as well, and the abundance of the future can be shared by everyone." Zuckerberg touts the potential upsides of superintelligence -- and Meta's place in advancing it -- but also warns that action is needed to maintain America's competitive edge and ensure humanity retains control over the technology. Here are five takeaways from Zuckerberg's essay. "Free or affordable access" to AI for all Zuckerberg envisions a future in which individuals have a personal agent designed to assist them with everyday needs related to their health, careers, finances and relationships. To encourage AI use, the tech billionaire said Meta will provide free versions of its tools to "billions of people" and seek to offer access to its AI technology for the "lowest price possible." Zuckerberg also committed to building a "fully private mode" for personal AI agents that will prevent Meta and other providers from seeing people's information. Jobs won't disappear Disputing widespread concerns that AI will displace workers, Zuckerberg said the technology will create new kinds of jobs that don't exist today and allow people to do more with less. In his essay, he referred to superintelligence as a "tool of invention" rather than of automation. "Company sizes may shrink -- just as they did in the transition from industrial giants to tech companies," Zuckerberg said. "But this doesn't mean fewer jobs overall. It implies a larger number of companies with fewer people each." Data center community investments Meta is among the so-called hyperscalers, along with rivals such as Alphabet, Amazon and Microsoft, rushing to build data centers as growing AI demand spurs a need for more computing power. Yet the push is drawing intense public opposition over environmental, economic and other concerns. Seeking to address such concerns, Zuckerberg said Monday that Meta will create an investment fund called the "Future is for Everyone Fund" to benefit communities where data centers are being built. The fund size will be $1 billion, according to the Wall Street Journal. Zuckerberg also committed to keeping electricity prices low, restoring local water supply and training workers to support its data center buildout. Meta's 2026 capital spending budget of $145 billion will be used "largely to build data centers," he noted. Releasing open-source models Zuckerberg argues in his essay that the U.S. should lower barriers that he said make it harder for U.S. open-source AI models to compete with those of foreign competitors. While there are risks to releasing models, allowing AI labs to keep them under lock and key as proprietary AI tools would be worse, he said. To that end, Zuckerberg said Meta will "resume releasing some open source models soon." According to the Wall Street Journal, Meta is planning to launch a new model with open weights -- numerical values that determine a model's behavior -- called Muse Glimmer, as well as an open-weight version of Muse Spark 1.2, its most advanced AI coding model. Working with the authorities Zuckerberg also offered policy suggestions for how AI companies and the government can work together to advance AI safely. Those include sharing " intermediate training checkpoints of new models" with the government -- allowing it to tackle potential security issues early on. AI labs should also work with law enforcement to prevent bad actors from misusing the models, he said.

Meta releases new AI model as Zuckerberg lays out vision

Washington (United States) (AFP) - Meta chief executive Mark Zuckerberg called Monday for lower barriers for open source AI, as his company introduced a new model that can run on a personal computer. The launch comes amid debate in the United States on whether access to the technology should be more restricted, and on local pushback against the buildout of data centers that power it. This was accompanied by a long essay outlining Zuckerberg's vision for artificial intelligence in which he stressed the need to remain competitive against countries like China, among other issues. Competitors OpenAI and Anthropic largely focus on closed AI models. Zuckerberg said in an Instagram video Monday that his company would open the weights for its latest AI model, enabling it to be downloaded by the public. Weights refer to calculations or rules governing how an AI system behaves. He said Meta would also release a new class of open source models, designed to run on a laptop. "Rather than centralizing superintelligence, we should distribute it widely and give every person the ability to direct it," Zuckerberg said in his essay. He took aim at other labs that are "focused on building AI for companies, governments or other institutions," arguing that this would tip the balance of power in favor of institutions instead of individuals. Zuckerberg added that foreign labs hold advantages over US ones that have to comply with restrictions on training data. "I do not believe restricting access to foreign open source models is an effective solution," he said. "Our goal should be for American open source models to be the best globally. This requires removing the hurdles that make it harder for American open source models to compete," he said. Amid growing public opposition to data center development in the United States, Zuckerberg added that Meta is launching a fund to directly support communities around its facilities.

Meta releases open-source Muse Glimmer model with 30B parameters

Meta Platforms Inc. today released Muse Glimmer, an open-source language model that can run on personal computers. The company also published a lengthy essay penned by Chief Executive Officer Mark Zuckerberg. The document discusses the risks of artificial intelligence, open-source model regulations and several related topics. Muse Glimmer features 30 billion parameters, which means that it would normally require about 55 gigabytes of RAM. Meta's engineers shrunk its footprint to under 20 gigabytes using various optimization methods. As a result, it can run on personal computers and Macs with a single consumer-grade graphics card. One of the optimization methods Meta used is called quantization. It compressed each of the model's weights, the configuration settings that determine how it processes data, into four bits. The company determined that the quantization introduced "minimal to no degradation on agentic tasks." Muse Glimmer generates prompt responses through a two-step process. First, it uses a less advanced "drafter" model to output an initial answer. It then verifies the accuracy of the drafter's answer, refines it and delivers the polished response to the user. That approach, which is known as speculative decoding, is faster than having Muse Spark generate prompt responses on its own. Meta trained the model on data generated by its flagship Muse Spark series of proprietary AI models. The company then refined Muse Glimmer through two additional training runs. In the first run, Meta enhanced the model's ability to tackle lengthy prompts and reasoning tasks. The second training session made Muse Glimmer better at powering AI agents. Some models stop running if they encounter an obstacle while processing a prompt. According to Meta, its engineers trained Muse Glimmer to retry tasks that it fails to complete on the first attempt. Users can adjust how much time and computing power Muse Glimmer spends on each task thanks to built-in "reasoning strength" settings. Meta tested the model across two dozen popular AI benchmarks. According to the company, Muse Glimmer outperformed the comparably-sized Gemma4-31B and Qwen3.6-27B across half the benchmarks. The evaluations in which the model won first place covered tasks such as online research, code generation and scientific chart analysis. The launch of Muse Glimmer comes more than a year after Meta released its last open-source AI model. In today's essay, Zuckerberg wrote that the company plans to resume releasing open-source models. He wrote that the next AI release will be "soon." The lengthy document also covers a range of other topics. It contains predictions about the economic benefits of AI, a lengthy discussion of the technology's risks and suggestions to policymakers. In particular, the essay calls on the U.S. government to reduce regulatory obstacles to open-source model development. According to Zuckerberg, Meta's board is adopting a governance structure that will enable it to define AI safety criteria. The company will use those criteria to evaluate each of its future models and determine whether they should be broadly released. The essay calls on other frontier AI developers to adopt similar measures. Additionally, Zuckerberg argues that AI labs should give the government access to new models while they're still being trained. According to the executive, such early access would make it easier to address AI-related cybersecurity risks.

Mark Zuckerberg lays out Meta's AI vision in a 6,500-word essay. 6 things to know

Mark Zuckerberg is making the case for Meta's approach to artificial intelligence. In a 6,500-word essay published Monday, Zuckerberg explained the company's thinking on AI, what it could mean for society and security, and how policymakers should approach calls for greater oversight of the industry. The Meta founder championed open-weight AI and announced a $1 billion fund to invest in communities where Meta operates data centers. Here are the biggest takeaways from the essay. Open-source is the future The unwritten thrust of the essay was to introduce Meta's new AI model, called Muse Glimmer, which will include open weights (letting users download the information that determines the model's behavior). While Zuckerberg does not mention it by name, the essay extols the benefits of open-weight (or open-source) models.

Mark Zuckerberg says the future will have an 'abundance of jobs' -- and predicts a wave of new careers like world builders and personal biologists | Fortune

"I do not understand why anyone who believes that AI will eliminate most jobs and much of humanity's relevance would rush to build that future," Zuckerberg wrote in an essay posted to Meta's website Monday. "The notion that AI is so dangerous that the only safe path is an extreme concentration of power seems inherently problematic. Historically, hoping that an absolute power will benevolently provide for humanity if sufficiently enlightened has not led to safe or positive outcomes." Zuckerberg, who is the world's sixth-richest person with a net worth of $211 billion, sees AI as a catalyst for an "abundance of jobs in the future," arguing that people will continue to invent new products and services -- and the jobs needed to bring them to life. He pointed to the fact that a generation ago, commonplace jobs like app developers, social media creators, electric vehicle technicians, and data center operators, barely existed. And moving forward, the 42-year-old expects a new set of careers to spawn up, such as one-person studio designers creating physical products, world builders building virtual experiences, and personal biologists formulating personalized treatments. Zuckerberg admits AI automation could bring a 'difficult period' for workers in the interim Despite Zuckerberg's largely optimistic tone, he acknowledged that the transition could still come with growing pains. If automation outpaces workers' ability to develop new skills, job displacement could create a "difficult period" for society. That disruption could be particularly painful without greater investment in training workers in career paths less vulnerable to automation, like in the skilled trades. "Unlike the concern about knowledge work displacement, there is a shortage of skilled tradespeople like carpenters, electricians, and construction workers to support the demand for infrastructure buildout," he said. Earlier this year, Meta committed $115 million into building out America's Workforce Academy -- a training program designed to prepare workers for data-center technician roles. The five-week program provides training at no cost, and graduates are guaranteed a job at one of Meta's data-center sites. Zuckerberg says higher education isn't preparing young people for jobs Preparing workers for an AI-driven economy isn't just about creating new career pathways -- it also raises broader questions about how people should be educated and trained in the first place. Zuckerberg, a Harvard dropout, has long argued that higher education can leave young people with significant debt without necessarily preparing them for the jobs they will eventually need to fill. "It would be one thing if [college] was just kind of like a social experience, but you started off neutral. The fact that it's not preparing you for the jobs that you need and you're kind of starting off in this big [financial] hole then I think that's not good," Zuckerberg said on Theo Von's This Past Weekend podcast last year. "There's going to have to be a reckoning...and people are going to have to figure out whether that makes sense. It's sort of been this taboo thing to say, 'Maybe not everyone needs to go to college,' and because there's a lot of jobs that don't require that...people are probably coming around to that opinion a little more now than maybe like 10 years ago," he added. But if Zuckerberg's vision of the future comes to fruition, AI could make traditional education itself far less necessary. "Everyone will have a personalized tutor and coach with a PhD in every subject and unlimited patience to help you learn anything you want," Zuckerberg wrote. "Students will have extra help in areas they need it that is currently only available to those whose parents can pay. Adults will have a superintelligent learning assistant that knows exactly how to teach you new job skills, new languages, new hobbies, or anything else you're interested in."

Meta open-sources Muse Glimmer for running AI agents locally

Meta has released Muse Glimmer, a new open-source AI model that can run on a single computer with one GPU for agent tasks such as scheduling and file management. The company said Muse Glimmer is a slimmed-down model based on its closed Spark 1.2 model and that the download is available for free for users to run on their own PCs. Meta said it will make the model's weights available on Hugging Face along with developer documentation. The company added that optimized integrations are planned for llama.cpp and other platforms so users can move from download to a working agent quickly. "We designed Muse Glimmer to balance capability against the memory and compute constraints of local hardware," Meta said. Meta said the model posts strong success rates on benchmarks including DeepSearch QA, MCP-Atlas and SWE-Bench, which tests code writing and debugging. The company also said the model supports tool use, multi-step reasoning, failure recovery, multimodal input and scaffold compatibility with OpenClaw and other agent orchestrators. Meta said Muse Glimmer was trained on data from more than 100 languages. The release adds a lighter local-running model to Meta's AI lineup as competition grows from models such as DeepSeek, which can also run on local machines and offers more open licensing. "Rather than centralizing superintelligence, we should distribute it widely and give every person the ability to direct it," Chief Executive Mark Zuckerberg said in an essay published alongside the release.

Meta releases scaled-down AI model consumers can use at home

Meta Platforms Inc. introduced a new AI model that's light enough to run on a single computer, allowing users to download and customize the technology. The new Muse Glimmer is a distilled version of the company's Muse Spark 1.2 model, designed with a focus on efficiency to minimize system requirements, Meta said on Monday. At 30 billion parameters, the Muse Glimmer is "small enough" to need only one graphics card to power its work, which will primarily involve agent-like artificial intelligence tasks such as schedule management and file organization, according to Meta. The weights, or values that help the AI system make decisions, for Muse Glimmer will be available on Hugging Face and Meta also intends to make the weights for a version of the more powerful Muse Spark available as well. The move to release these AI models under a permissive license mirrors competition in China, where startups like DeepSeek and Moonshot, along with internet giant Alibaba Group Holding Ltd., have adopted the practice to attract the largest number of users. Meta is in a global race to deliver AI breakthroughs and benefits that would eventually pay off lavish spending on hardware and data centers to develop the technology. Along with U.S. peers Amazon.com Inc., Alphabet Inc. and Microsoft Corp., Meta has committed more than $2 trillion to expanding its capabilities, and developing AI models and tools that entice people to use it over those rivals will help the company secure long-term business. It's developing plans for a cloud infrastructure business that will sell access to AI computing power and models, along the lines of Amazon Web Services. Menlo Park, California-based Meta also said on Monday that it intends to set up a $1 billion fund to invest in US communities where Meta owns and operates data centers. The runaway spending by hyperscalers like Meta and Google has been met with resistance in many locations across the US where the companies are setting up resource-hungry operations. -With assistance from Kurt Wagner. More stories like this are available on bloomberg.com

Mark Zuckerberg Thinks the Future Won't Have Fewer Jobs. It Will Have More Companies With Fewer People

Mark Zuckerberg has a giant platform to speak from, and today, he used it. In a new 6,500-word blogpost, the Meta CEO called out people who are doomy about the future of AI. We're "fortunate to live at an incredible moment in history," he wrote, arguing for AI in the form of next-generation superintelligence. He went on to promise that Meta will lead this revolution by releasing "open source models, and delivering them to billions of people around the world," expanding on a position he stated recently. But the biggest idea in his essay is his picture of our "incredible" workplace future, which includes just as much employment as there is today. It's a strong counterargument to the many experts who predict mass AI unemployment. According to Zuck, over the next couple of years "people will be able to use superintelligence beyond human capacity to create and discover extraordinary new things," and to "build new businesses, express new ideas." As long as we're careful about managing the "opportunities and challenges" from a firm philosophical viewpoint, he thinks AI will be a "tool that empowers everyone," rather than a tool that puts entry-level workers, financial experts, mathematicians, mid-level managers and, eventually, everyone out of work. What will happen instead, Zuck thinks, is that the world will undergo an employment paradigm shift, and traditional models for how a company works will be radically upended as AI takes on a more central role. For starters, he thinks "company sizes may shrink," just like they did "in the transition from industrial giants to tech companies," where thousands of, say, radio factory workers were replaced by precision machines and robots making integrated circuitry. But this time it won't result in fewer jobs, Zuck said. Instead it "implies a larger number of companies with fewer people each." He expects soon "we will start seeing small numbers of people with personal superintelligence agents able to run companies at significant scale." His assertion is appealing. It resonates with many pro-AI arguments that say that what these new tools are doing is freeing up workers from mundane tasks so they can deliver a human touch on more complex or more creative business matters. Zuck's taking this argument to its extreme, picturing a world where small businesses are less associated with early-stage startups or small, low-revenue mom-and-pop outfits, and are more akin to micro-staffed tech giants pulling in large-scale revenues. Think of it as today's influencer, freelance worker, or solopreneur model amplified a thousand times by powerful AI tech. Zuck explicitly imagines a future in which "small businesses will continue to be the backbone of the economy, but each small business will be able to have a much larger impact." To say that this will be a dramatic upset to millions of people is a staggering understatement. Zuck agreed that "people will have to adapt, and this will be challenging," with the word "challenging" doing a lot of work here, since it neatly skirts questions of social unrest or government failures that would inevitably go along with this kind of dramatic shift. Zuck also imagines a technological solution to these issues, picturing that everyone will have a "a personal agent that is superintelligent at teaching us new skills and helping us adapt to change, the smoother this will be." Essentially, AI will help us learn to adapt and embrace the new AI-powered economy. Even if this view of the future is only partially correct, it means the business world is set for dramatic upsets to pretty much every workplace norm. This is exciting, but it should also worry savvy leaders who fret that they may have to scale down staffing, or dramatically scale-up AI training and change company culture so that it can adapt, in an agile way, to new AI innovations. This will be a complex, and potentially expensive part of a good five-year business plan in the AI era. Get 1 Smart Business Story delivered straight to your inbox when you subscribe to Inc.'s free daily newsletter.

Zuckerberg argues most common AI concerns overblown

Meta CEO Mark Zuckerberg published a Monday blog post on the company's website touting artificial superintelligence for its accessibility, personalized features and ability to serve as an economic power tool amid concerns about humans' dwindling control over large language models. Artificial intelligence (AI) is the current technology that mimics human tasks like writing or driving, while Artificial Superintelligence (ASI) is a hypothetical future stage where a computer system surpasses the total cognitive capability and creativity of the smartest human minds across every domain. "I do not understand why anyone who believes that AI will eliminate most jobs and much of humanity's relevance would rush to build that future. The notion that AI is so dangerous that the only safe path is an extreme concentration of power seems inherently problematic," Zuckerberg wrote in the Monday blog post, a piece of which was originally published as an excerpt for an op-ed in the Wall Street Journal. "Historically, hoping that an absolute power will benevolently provide for humanity if sufficiently enlightened has not led to safe or positive outcomes," he added, noting that previous developments that drew concerns for the future of humanity advanced "prosperity, health and freedom." His comments come in contrast to Geoffrey Hinton, known as the "godfather" of AI, who has repeatedly warned that humans run the risk of being outsmarted by the next generation of large language models. He, alongside other coders from OpenAI and Google DeepMind, have urged government officials to put regulations in place to prevent fast-paced development before it reaches levels that endanger society. In his blog post, Zuckerberg said he supports government policy to ensure a "positive future" for new developments but opposes current guardrails in the U.S. Last week, a Meta AI model went rogue in the testing phase and attacked another company. Several similar attacks launched by models without human oversight have been reported by Anthropic and OpenAI. Still, Zuckerberg says a future with AI will ultimately be more beneficial than life would be without it. "People fear that automation will outpace individuals' capability growth, leading to job displacement followed by a difficult period as people learn new jobs. But there is no rule that AI must increase automation faster than it increases individuals' capabilities or demand for new skills. Recent statistics suggest it may be more likely that individuals' capability growth could match or outpace automation, in which case people will gain the ability to do many new things before their current jobs change," Zuckerberg wrote regarding AI's impact on employment. "This would lead to a healthy balance and potentially even job growth. Which outcome we get depends on the balance in progress between automation on one side and individual empowerment and invention on the other," he added. Zuckerberg says AI is a "tool" for the future. "There are several reasons to be optimistic that there will be an abundance of jobs in the future. No matter how intelligent AI becomes, there will always be a finite amount of compute and therefore an opportunity cost for how we use it," Zuckerberg wrote. "If people can use AI to invent incredibly valuable new things, then it will make more sense to allocate it towards that rather than automating existing jobs. The more superintelligence serves as a tool of invention, the more likely that individual capability outpaces automation and the future is better for people," he added. The Meta CEO says artificial superintelligence could help combat government tyranny through personal access to large language models, advance science and expand communities through the "most profound technological advance we will see in our lifetimes." "As we get closer to this, it is increasingly important to have a clear philosophy and values for how superintelligence will benefit humanity. Meta is committed to building with the principles of individual empowerment as the source of prosperity, invention as AI's purpose, and a balance of power favoring people as the foundation for addressing safety risks," Zuckerberg wrote. "If these values lead the way, then I am optimistic that the coming decades will be some of the most amazing in history. The arc of human civilization has bent towards putting more power in people's hands to live and shape the world in the ways we believe are best. Superintelligence holds the promise of giving everyone that power, and building a positive future for everyone," he added.

Meta launches new Muse Glimmer AI model

Aug 10 (Reuters) - Meta META.O CEO Mark Zuckerberg called for lower U.S. barriers for open-source AI models to compete with Chinese rivals as the social media giant released a new open-weight model on Monday. The new model, Muse Glimmer, is much smaller than leading AI models from rivals and is instead designed for agentic tasks and can run on a Mac or PC with a single graphics card, aiming to tap demand for AI systems that run directly on people's devices. It comes as the social media giant seeks to strengthen its position after forming a costly, new superintelligence team last year to propel itself back into the high-stakes AI race. Zuckerberg's statement also marks the latest show of support for open-weight AI, which is gaining traction as businesses grow wary of ballooning AI bills and worry about recent cybersecurity incidents involving models from Anthropic, OpenAI and Meta. Hugging Face, the AI coding collaboration site that was hacked by a rogue OpenAI model, said last month that it used a Chinese open-weight model to defend against the attack because closed-source models have restrictions on use for cybersecurity work. Open-weight models are typically cheaper than leading models from so-called frontier labs such as OpenAI and Anthropic. Open-weight models also come with publicly accessible core components for easy customization, unlike closed models that companies keep fully under their control. Policy rethink needed to propel open-weight Zuckerberg said in a statement that the U.S. needed to rethink policies if domestic firms were to lead in open-weight models. Chinese startups are leading the race for open-weight models, with Moonshot's Kimi K3, alongside Alibaba's Qwen3.8-Max and DeepSeek's V4-Flash, delivering performance that rivals top systems by U.S. AI labs. By contrast, the leading models of U.S. developers OpenAI, Anthropic and Alphabet's GOOGL.O Google are closed source. "Foreign labs currently hold several advantages here since American labs have to comply with many additional restrictions on training data," Zuckerberg said, referring to open-source models. "US policy must reduce this additional friction if we want American open source models to lead over time," Zuckerberg said, adding that restricting access to foreign open-source models was not an effective solution. U.S. President Donald Trump's administration told AI developers earlier this month that it will not put open-weight AI models through voluntary safety tests, according to two sources familiar with the discussions. In his statement, Zuckerberg also advocated for AI model distillation, or using a powerful AI system to train a smaller model. He said that Meta would implement a governance structure to give its independent directors the power to approve the safety criteria for releasing models. (Reporting by Ananya Palyekar and Shubham Kalia in Bengaluru; Editing by Mrigank Dhaniwala and Devika Syamnath)

Meta's Muse Glimmer AI Model Unveiled: All You Need to Know

Muse Glimmer can operate on a Mac or PC with a single consumer GPU Meta has introduced Muse Glimmer, a new AI model developed by its Meta Superintelligence Labs. The Muse Glimmer features 30 billion parameters and is optimised for always-on local agent workflows. Meta says it is suitable to run on a Mac or PC equipped with a single consumer GPU. At full precision, Muse Glimmer would require more than 55GB of memory. The Muse Glimmer is designed to handle functions including agentic task completion, multi-step reasoning, and multi-model input and reasoning. Meta Launches Muse Glimmer In a newsroom post on Monday, Meta has announced Muse Glimmer. This latest AI model from Meta Superintelligence Labs has the model weights under a permissive Apache 2.0 license. "Muse Glimmer is a 30-billion-parameter model optimised for always-on local agent workflows", said Meta. The company says the model can work on a Mac or PC with a single consumer GPU. This would enable use cases including local agents and function calling, local coding, and LLM-as-a-judge evaluation. Meta claims that the Muse Glimmer offers strong performance on key agentic use cases and benchmarks compared with leading models in the same size category. The company says many AI applications still depend on cloud infrastructure and network access, and Muse Glimmer is designed to address this limitation and is optimised for local use cases. It is claimed to allow Running models locally with or without an internet connection. The 30-billion-parameter model would require more than 55GB of memory at full precision. Meta uses 4-bit quantisation to reduce the model size to under 20GB, and this lets it run on a 24GB or 32GB memory setup. Meta has also shared a direct comparison of the Muse Glimmer with Google's Gemma4-31b and Alibaba's Qwen3.6-27B. The results show that Muse Glimmer dominates both models in 12 tests. According to Meta, Muse Glimmer is built for end-to-end agentic task completion, reliable tool use, multimodal reasoning, error diagnosis and tool retries, multi-step reasoning, among others. It is also trained on data from over 100 languages. Users can download the weights from Hugging Face. Meta says the Muse Glimmer will be available in the coming days from apps like Ollama, LM Studio, and Unsloth. The tech giant confirmed that it is working with AMD, Arm, Dell, Intel, and Nvidia to optimise performance across devices. Meta confirmed that the optimised integrations for llama.cpp, MLX and ExecuTorch will be available in the coming days.

ETtech Explainer: Meta's stop-start strategy for its open source AI models

Glimmer, a 30-billion-parameter model, is much smaller than leading AI models from rivals and is designed for agentic tasks. It can run on a Mac or PC with a single graphics card, aiming to tap into demand for artificial intelligence (AI) systems that run ⁠directly ⁠on people's devices. Mark Zuckerberg's Meta is back on the open source bandwagon with its new model Muse Glimmer, which was released this week.Glimmer, a 30-billion-parameter model, is much smaller than leading AI models from rivals and is designed for agentic tasks. It can run on a Mac or PC with a single graphics card, aiming to tap into demand for artificial intelligence (AI) systems that run ⁠directly ⁠on people's devices.The tech giant's approach to its models

Meta's Muse Glimmer Runs on a Laptop -- and Will Keep Your Sensitive Files Out of the Cloud

Thanks to its smaller size, Muse Glimmer is less powerful than top-tier AI models but can be downloaded and run locally, offering more peace of mind for privacy-conscious users who don't want their data going to the cloud. As companies shift from "tokenmaxxing" to "valuemaxxing," local models also offer a business advantage, as they are free to use and have no usage caps. Local models like this one are not new, but as models and hardware have improved, they are more powerful than they used to be. Meta had a line of small open-source models called "llama," but updates to those models slowed early last year. Google and OpenAI have their own versions of open-source models, but they are much less powerful than their closed-source counterparts.

Meta's Muse Glimmer 30B Outperforms Qwen 3.627B in Benchmarks

Meta's latest release, Muse Glimmer 30B, marks a significant development in the open AI landscape. With a dense architecture using all 30 billion parameters during inference, the model is optimized for tasks requiring multi-step reasoning and long-term decision-making. Sam Witteveen explores how Muse Glimmer, licensed under Apache 2.0, is designed for local deployment, with quantized versions allowing efficient operation on GPUs like the Nvidia 3090 and AMD 9700. This focus on accessibility ensures that researchers and developers can use its capabilities without relying on large-scale infrastructure. Discover how Muse Glimmer's speculative decoding feature enhances real-time performance by reducing latency, making it suitable for applications demanding speed and responsiveness. You'll also gain insight into the advanced training methodologies Meta employed, including distillation and reinforcement learning, to refine the model's accuracy and reliability. Whether you're interested in its benchmark performance or its potential to support diverse AI applications, this explainer provides a detailed look at what sets Muse Glimmer apart. What Distinguishes Muse Glimmer? Muse Glimmer stands out as a dense model, meaning all its parameters are actively utilized during inference. This contrasts with mixture-of-experts models, which selectively activate subsets of parameters. With its 30 billion parameters, Muse Glimmer excels in tasks requiring multi-step reasoning, effective tool usage and long-term decision-making. These capabilities make it particularly well-suited for applications involving complex problem-solving and sustained analytical tasks. When compared to Qwen 3.627B, Muse Glimmer demonstrates a clear competitive edge across multiple benchmarks. Its robust architecture and performance solidify its position as a high-performance model, capable of addressing a wide range of challenges in AI research and practical applications. Innovative Training Methodologies Meta employed advanced methodologies to develop Muse Glimmer, combining distillation and reinforcement learning to refine the model after its initial training phase. Unlike traditional pre-training approaches that rely heavily on raw internet data, Muse Glimmer was pre-trained using outputs from larger Muse Spark models. This strategy ensures the model benefits from high-quality, curated data, resulting in improved accuracy, reliability and overall performance. These innovations reflect Meta's commitment to pushing the boundaries of AI training techniques. By using curated data and advanced refinement processes, Muse Glimmer achieves a level of precision that enhances its usability across diverse applications. Here are more detailed guides and articles that you may find helpful on Meta AI. Optimized for Local Deployment Muse Glimmer is specifically designed to support local usability, making it accessible to developers and researchers working with consumer-grade hardware. A quantized 4-bit version of the model ensures efficient operation on GPUs such as the Nvidia 3090, 4090 and AMD 9700, as well as devices with 24GB or 32GB of memory. This optimization allows the model to run smoothly while maintaining sufficient memory for KV cache management, making sure reliable performance during intensive tasks. Additionally, Muse Glimmer is compatible with high-end laptops like the MacBook Pro (64GB memory), further broadening its accessibility. This focus on local deployment enables developers to experiment with and implement AI solutions without requiring expensive, large-scale infrastructure. Key Technical Features Muse Glimmer incorporates several advanced features that enhance its performance and usability, making it a versatile tool for a wide range of applications: * Speculative Decoding: This feature significantly improves decoding efficiency, reducing latency during inference. It ensures the model is suitable for real-time applications, where speed and responsiveness are critical. * GPU Optimization: The model's architecture is fine-tuned to maximize GPU performance, allowing it to handle demanding tasks without compromising on speed or accuracy. * Scalability: Muse Glimmer is designed to support both research-focused and practical applications, making it a flexible solution for developers and researchers alike. These features collectively position Muse Glimmer as a powerful and adaptable AI tool, capable of addressing the diverse needs of the AI community. Meta's Strategic Vision The release of Muse Glimmer aligns with Meta's broader vision of contributing to the AI community by offering open, high-performance models. This initiative builds on the success of previous releases like Llama and reflects Meta's ongoing commitment to fostering innovation and collaboration. By prioritizing accessibility and performance, Meta aims to empower developers and researchers with innovative tools that drive progress in AI development. Meta has already announced future releases, including Muse Spark 1.2 and Muse Code, which are expected to further expand the capabilities of its AI offerings. These upcoming models will likely enhance the versatility of AI tools for agents and local deployment, reinforcing Meta's role as a leader in the field. Impact on the AI Community By making Muse Glimmer's open weights available on the Hugging Face platform, Meta has demonstrated its dedication to transparency and collaboration. This decision enables the AI community to build upon the model's capabilities, fostering innovation and driving progress in AI research and development. The release has sparked anticipation for future advancements, including comparisons with upcoming models like Qwen 3.827B. Muse Glimmer is poised to influence the trajectory of AI research, encouraging collaboration and setting new standards for performance and accessibility. Its dense architecture, innovative training methodologies and local deployment capabilities make it a valuable resource for researchers and developers seeking to push the boundaries of artificial intelligence. Media Credit: Sam Witteveen Disclosure: Some of our articles include affiliate links. If you buy something through one of these links, Geeky Gadgets may earn an affiliate commission. Learn about our Disclosure Policy.

Meta Fires Back At China's 15-Week AI Token Dominance With Muse Glimmer, Which Can Fit Inside A Single GPU And Uses An Innovative Technique To Speed Up Responses

Meta is returning to the open-weight AI model category - a subset of large language models (LLMs) that it founded but then largely abandoned in a bout of misplaced priorities - with a loud and fairly sonorous bang, courtesy of the just-released Muse Glimmer open-weight model that employs creative tricks to make sure the model fits inside a single consumer GPU, and responds to queries in a lightning-fast manner. Meta used distillation to ensure that Muse Glimmer fits inside a single consumer GPU, while a tiny companion model speeds up response times Meta has just released the Muse Glimmer, a 30-billion-parameter open-weight AI model that is meant to compete with the likes of Google's Gemma 4 and Alibaba's Qwen 3.6. For the benefit of those who might not be aware, weights tell a model how much importance it should accord to any given concept, and represent the entire breadth of that model's knowledge base. Meta's Muse Glimmer is distinctive for two major reasons. First, to run a 30-billion-parameter at full precision (fp16), you need around 60GB (30x10⁹x2) of memory for model weights alone. However, Meta used quantization - essentially distillation, where a larger model (Muse Spark) trains a smaller one, imbuing it with many of its capabilities - to compress it to reduce the memory requirements for model weights down to 20GB. Coupled with 2GB to 4GB required for KV cache, these requirements are modest enough that a 24GB or 32GB consumer graphic card can easily run Muse Glimmer. According to Meta, this model compression via distillation produces no noticeable impact on performance. Second, to speed up token generation (response times), the Muse Glimmer uses a tiny companion model - called the DFlash drafter model - that predicts whole chunks of text, allowing the Muse Glimmer to then check the answer in one go, keeping correct text responses while discarding erroneous ones. Since checking an answer is much faster than generating one, the arrangement allows for 3.1x faster responses on an RTX 5090 card, 1.8x on an M5 Max, and 1.5x on an M4 Max, as per the data disclosed by Meta. Of course, Meta's Muse Glimmer could not have come at a more opportune time for the Western open-weight AI model landscape. After all, according to the data from OpenRouter, Chinese LLMs recently crossed the 34.25 trillion weekly tokens level for the first time, with DeepSeek's V4 Flash recording an astounding 570 percent week-over-week growth. According to the latest OpenRouter dataset that pertains to the week that began on August 03, global AI model usage hit 69 trillion tokens, up 21.48 percent week-on-week. Chinese models accounted for 34.25 trillion of that total, while US models contributed just 9.17 trillion tokens to that cumulative figure. Critically, it was the fifteenth consecutive week that Chinese models have led the global count. Consequently, we'll be paying particular attention to next week's figures to see if Meta's Muse Glimmer is having a noticeable impact on the relentless ascendancy of Chinese open-weight AI models. Follow Wccftech on Google to get more of our news coverage in your feeds.

Zuckerberg's $145B Meta AI Bet Is Getting Serious - Meta Platforms (NASDAQ:META)

Zuckerberg Says AI Should Belong to Everyone. Meta Is Spending $145 Billion to Make It Happen. Mark Zuckerberg has a simple answer to one of the biggest questions in AI: Who should get superintelligence? Everyone. The vision comes with an enormous infrastructure bill. Meta -- which owns Facebook, Instagram and WhatsApp -- expects to spend $130 billion to $145 billion on capital expenditures in 2026. That includes investments in data centers and other infrastructure. The company spent $31.1 billion on capital expenditures in the second quarter alone. Not all of that spending is specifically for Zuckerberg's vision of personal superintelligence. But the scale of the investment shows how seriously Meta is preparing for an AI future in which increasingly powerful models become part of everyday life. Zuckerberg Wants AI in Everyone's Hands Zuckerberg's argument is broader than simply making another chatbot available. His central concern is that if increasingly powerful AI is controlled by only a small number of institutions, it could concentrate too much power in too few hands. Meta's strategy is to distribute that technology through products used by billions of people. That approach became more tangible Monday when Meta released Muse Glimmer, a smaller AI model designed to run on personal computers using a single graphics card. Zuckerberg also said a more advanced Muse Spark 1.2 model is coming soon. Getting there will not be cheap. The $145 Billion Question Zuckerberg is effectively betting that today's infrastructure spending will create tomorrow's AI platform. And Meta has an advantage that many AI rivals lack: distribution. Its Family of Apps reached 3.6 billion daily active people in June, giving Meta an enormous audience to which it can introduce AI products. That makes Zuckerberg's argument about access more than a philosophical statement. It is also a business strategy. Meta wants to build the models, spend heavily on the infrastructure behind them and put those models in front of billions of people. For META investors, the question is no longer whether Zuckerberg is willing to spend heavily on AI. He clearly is. The bigger question is whether making superintelligence affordable can generate enough value to justify the extraordinary cost of building it. But while Zuckerberg sees personal AI as a tool to boost careers, businesses, education and creative work, Meta's own upheaval highlights the darker side of the AI boom: tech companies are increasingly cutting jobs as AI takes on work once done by employees. In May, Meta laid off more than 8,000 workers, or about 10% of its workforce, underscoring how AI is reshaping the jobs it was supposed to help. Image via Shutterstock Market News and Data brought to you by Benzinga APIs To add Benzinga News as your preferred source on Google, click here.

Meta launches new AI model as Zuckerberg lays out vision for open-weight AI

Meta Platforms CEO Mark Zuckerberg called for lower U.S. barriers for open-source AI to better compete with Chinese rivals as the social media giant released a new open-weight model on Monday and said more would follow soon. The new model, Muse Glimmer, is smaller than leading AI models from rivals and is instead designed to run agentic tasks on a Mac or PC with a single graphics card, aiming to tap demand for AI systems that run directly on people's devices. "In the coming weeks, we are also going to open the weights for MuseSpark 1.2, our latest foundation model and one of the leading models in the world. And we've got even bigger models that are coming soon too," Zuckerberg said in a video post accompanying his 14-page essay titled "The Future is for Everyone" in which he championed spreading AI rather than leaving it in the hands of a few.

Zuckerberg's Essay Says the Future of AI Is Personal Superintelligence | PYMNTS.com

In an essay titled "The Future is for Everyone" published Monday (Aug. 10), the Meta CEO argued that superintelligence should not sit inside a handful of labs, governments or companies. It should sit in everyone's hands, he wrote, deployed the way personal computers and the internet eventually were. The core claim is specific. Zuckerberg proposed "a philosophy based on individual empowerment as the source of prosperity, invention as the primary purpose of superintelligence, and balance of power as the foundation of safety." He is betting Meta's AI strategy on that framing. Invention Over Automation Is the Operating Thesis "Invention, not automation, will be the greatest contribution of superintelligence," Zuckerberg wrote. Early AI answered questions and handled routine tasks. He expects the next wave to discover new knowledge instead, from new drugs to new ways of running a business. That distinction shapes how he wants the technology distributed. "Rather than centralizing superintelligence, we should distribute it widely and give every person the ability to direct it," he wrote. Some in the industry argue that superintelligence itself, or a small set of experts who control it, should decide what's best for humanity. He disagreed, citing the history of democracy and economics as proof that no single actor can define the best life for everyone. He offered specific mechanisms. Meta plans a personal agent covering relationships, health, finances and hobbies, paired with a fully private mode the company says even it cannot access. It also plans a dynamic auction system for compute, designed to hold prices down while directing capacity toward whatever people value most. "My agent flags interesting information and helps me prototype ideas," Zuckerberg wrote, describing his own use of an early version. Zuckerberg's Safety Case Rests on Distributed Power Zuckerberg's safety argument breaks from how most labs talk about alignment. He said engineering a single benevolent superintelligence is the wrong target. Values differ too much across people for one system to serve everyone's interests at once. He illustrated the alternative with a thought experiment. If one person alone had a superintelligent lawyer, they would win cases regardless of merit, he wrote, calling that outcome a worse society. If everyone had one, "justice would be carried out much more fairly and efficiently than it is today." He applied the same logic to cybersecurity, arguing that widely distributed defensive tools would harden every system rather than concentrate power in whoever gets there first. He extended it to markets, warning that a single business holding superintelligence would out-compete everyone else and shrink the field of competitors. "There is no such thing as a singular benevolent superintelligence," he wrote. "Meta is the company primarily focused on building personal superintelligence for everyone," he added, arguing that most other labs build AI for companies, governments or institutions instead. What Comes Next for Companies and Governments Company sizes may shrink the way they did when industrial giants gave way to tech companies, Zuckerberg wrote, but that implies more companies with fewer people each, not fewer jobs overall. He expects small teams running personal superintelligence agents to operate at a scale that currently requires far more headcount. On security, he made his most specific policy ask. Zuckerberg proposed that "frontier AI labs should share intermediate training checkpoints of new models for government use and review rather than waiting until training has completed," giving government "early access to the most powerful models and an army of capable engineers to identify and patch security issues." He framed this to harden infrastructure without slowing releases or delaying public access.

Meta AI Releases Muse Glimmer 30B for 24GB VRAM Systems

Meta AI's Muse Glimmer 30B is a 30-billion-parameter model designed for local AI deployment, with features tailored to tasks such as autonomous agent creation and multi-step planning. It supports a 128k token context window and 4-bit quantization, making it accessible for systems with as little as 24GB of VRAM. According to World of AI, the model's strengths include workflow automation and token efficiency, though it faces challenges in areas like coding and system-level operations. Explore how Muse Glimmer 30B handles practical applications such as task automation, project management and creative content generation. Gain insight into its hardware requirements, specific use cases and how it compares to other open source models like Qwen 3.6 27B. This overview equips you with the details needed to assess its suitability for your projects. Core Features of Muse Glimmer 30B Muse Glimmer 30B is purpose-built for agentic workflows and long-horizon planning, making it a strong contender in automation and tool integration tasks. Its multimodal capabilities allow it to process diverse input types, enhancing its adaptability across a wide range of applications. Additionally, the model is optimized for token efficiency, reducing computational demands while maintaining high-quality outputs. Released under the Apache 2.0 license, it is freely available for open source use, encouraging collaboration and innovation within the AI community. Key specifications include: * 24GB VRAM requirement for optimal performance, making it accessible to users with mid-range hardware. * Support for 4-bit quantization, allowing efficient local deployment on lower-end GPUs. * A context window of 128k tokens, which, while sufficient for many tasks, may limit its utility in scenarios requiring extensive context retention. These features position Muse Glimmer 30B as a versatile tool for users seeking to deploy advanced AI models locally without relying on cloud-based solutions. Performance Insights Muse Glimmer 30B demonstrates exceptional performance in tasks requiring multi-step planning, error recovery and workflow automation. It ranks 23rd on the World of AI Benchmark and achieves an intelligent index score of 35, outperforming some larger models in specific scenarios. These strengths make it particularly effective for creating autonomous agents and managing complex workflows. However, the model is not without its drawbacks. It exhibits a high hallucination rate of 82% in knowledge-based tasks, which can undermine its reliability in certain applications. Additionally, its performance in coding and operating system-level tasks falls short compared to competitors like Qwen 3.6 27B, limiting its appeal for developers and technical users who require precision and depth in these areas. Here are more guides from our previous articles and guides related to Meta AI that you may find helpful. Hardware Requirements and Accessibility Muse Glimmer 30B is designed with local deployment in mind, making it accessible to users with mid-range hardware setups. For optimal performance, the model requires systems equipped with Nvidia GPUs, such as the RTX 3090 or higher. Its support for 4-bit quantization significantly reduces memory usage, allowing it to run on lower-end GPUs while maintaining efficiency. However, users with higher-end hardware will benefit from faster token generation speeds and improved overall performance, making it a better fit for demanding workflows. This hardware flexibility ensures that Muse Glimmer 30B can cater to a broad range of users, from hobbyists to professionals, who seek to use advanced AI capabilities without the need for extensive cloud infrastructure. Comparing Muse Glimmer 30B and Qwen 3.6 27B When comparing Muse Glimmer 30B to Qwen 3.6 27B, each model exhibits distinct strengths tailored to different user needs. Qwen excels in coding, reasoning and operating system-level tasks, making it an ideal choice for developers and system administrators. On the other hand, Muse Glimmer 30B outperforms Qwen in building autonomous agents, handling complex workflows, and executing agentic tasks. For users with diverse requirements, deploying both models locally could provide a more comprehensive solution, using the strengths of each to address a wider range of tasks effectively. Real-World Applications Muse Glimmer 30B is well-suited for a variety of practical applications, including: * Developing local agents for task automation and tool coordination, streamlining repetitive processes. * Managing workflows such as calendar scheduling, project management, and other organizational tasks. * Generating creative outputs, including SVG designs, simple web development projects, and other multimedia content. While the model performs adequately in creative and coding tasks, its capabilities in these areas are not new. Users with advanced coding or knowledge-based requirements may find Qwen 3.6 27B to be a more suitable option. Anticipated Developments in Open source AI The future of open source AI is poised for significant advancements, with Meta AI's Muse Spark 1.2 model expected to address some of the limitations of Muse Glimmer 30B. Improvements in coding and knowledge-based tasks could make it a more versatile tool for a broader range of applications. Similarly, the anticipated release of Qwen 3.8 is likely to further raise the bar, intensifying competition in the open source AI space. These developments highlight the rapid pace of innovation in AI technologies and their growing impact on local deployment, offering users increasingly powerful tools to meet their evolving needs. Contributions to the Open source Ecosystem Muse Glimmer 30B is a valuable addition to the open source AI ecosystem. By allowing the local deployment of advanced AI models, it reduces reliance on cloud-based solutions, offering users greater privacy and control over their data. Its release under the Apache 2.0 license fosters collaboration and innovation, paving the way for new applications and use cases. As the open source AI landscape continues to evolve, models like Muse Glimmer 30B play a pivotal role in shaping the future of autonomous systems and local AI deployment. By addressing its current limitations and building on its strengths, Muse Glimmer 30B has the potential to become an even more integral tool for users across various domains. Media Credit: WorldofAI Disclosure: Some of our articles include affiliate links. If you buy something through one of these links, Geeky Gadgets may earn an affiliate commission. Learn about our Disclosure Policy.

Meta introduces Muse Glimmer 30B open-weight model for local agent workflows

Meta Superintelligence Labs has introduced Muse Glimmer, a 30-billion-parameter open-weight model designed for always-on local agent workflows. The model is released under the Apache 2.0 license and is designed to run on a Mac or PC with a single consumer GPU. Muse Glimmer is aimed at local agents, function calling, coding, and LLM-as-a-judge evaluation. It is designed for AI workflows that can run locally without depending on cloud infrastructure or network access. How Muse Glimmer was trained Agents that manage schedules, draft messages, organize files, and adapt to how users work require access to personal context. Meta designed Muse Glimmer to combine long-horizon execution, precise tool calling, multimodal understanding, long-context memory, and instruction following. The model was designed to balance these capabilities with the memory and compute constraints of local hardware. Meta used a compact architecture, a distillation approach that transfers agentic reasoning from a much larger teacher model, and inference optimizations including quantization. The training process was divided into three stages: * Pre-training: Muse Glimmer was trained on Muse Spark's outputs using logit distillation, with a similar data mix as the teacher. * Mid-training: The model was trained on longer-context and more agent-focused data with richer reasoning traces, along with organic data. * Post-training: Meta combined supervised fine-tuning with on-policy distillation and reinforcement learning across general, reasoning, coding, and agentic domains. Meta evaluated Muse Glimmer under the standards of its Advanced AI Scaling Framework and assessed it for open-weight release across the relevant categories. Built for agentic workflows Muse Glimmer is trained and evaluated across capabilities required for agentic tasks. These include: * End-to-end task completion: The model is evaluated on DeepSearch QA, MCP-Atlas, τ-Bench, and SWE-Bench. These benchmarks measure its ability to work within scaffolds, write and debug code, and resolve multi-turn requests from start to finish. * Tool use: It handles a range of function calls and invokes tools using precise schemas throughout extended workflows. * Multi-step reasoning: The model chains reasoning over long horizons and sustains plans across complex, extended workflows. * Failure recovery: When a tool call fails or produces an unexpected result, the model is trained to diagnose the error and retry rather than halt. * Multimodal input and reasoning: A dedicated perception encoder allows the model to accept interleaved text and images and interpret screenshots, charts, and documents alongside conversations. * Scaffold compatibility: Muse Glimmer works with OpenClaw and other agentic orchestration patterns. * Controllable effort: The model supports different reasoning strengths to select a balance between quality and speed. * Multilingual: Muse Glimmer is trained on data covering more than 100 languages. Performance Meta evaluated Muse Glimmer across a broad range of benchmarks covering the capabilities required for autonomous agent behavior. The model was compared with Gemma4-31B and Qwen3.6-27B and, according to Meta, performs strongly for its size class on several widely used LLM benchmarks. Optimized for local deployment At full precision, a 30-billion-parameter model would require more than 55GB of memory. Meta uses quantization to compress the model weights to approximately 4-bit precision, reducing the language model to under 20GB. This leaves enough memory for the model's KV cache, the perception encoder for image understanding, and the speculative decoding drafter to run simultaneously within a 24GB or 32GB memory configuration. Meta says it validated the compression with minimal to no degradation on agentic tasks. Faster generation through speculative decoding Muse Glimmer uses a lightweight drafter model based on DFlash. The small companion network proposes blocks of tokens at once, after which the main model verifies the proposals in parallel, accepting correct tokens and correcting incorrect ones. Meta says this allows Muse Glimmer to generate text significantly faster than standard token-by-token generation while producing identical output quality. Quantized versions of the drafter are included to reduce additional memory overhead. Meta measured the K-Quant-17GB model alongside the quantized DFlash drafter on the MacBook M4 Max, MacBook M5 Max, and NVIDIA RTX 5090. Meta says the model is fast enough for fluid conversation and real-time agent interaction while running entirely on the device. Availability and deployment Muse Glimmer is available now as open model weights through Hugging Face under the Apache 2.0 license. Meta has also released developer documentation and resources for building and running agents, including guidance for setting up custom scaffolds. The model can be used across several local, edge, serving, and AI platforms: * Local platforms: Ollama, LM Studio, and Unsloth * Edge frameworks: llama.cpp, ExecuTorch, and MLX * Serving frameworks: vLLM and SGLang * AI platforms: Together AI, Fireworks AI, and OpenRouter Integrations with llama.cpp, MLX, and ExecuTorch are expected in the coming days. Developers can also customize Muse Glimmer using PyTorch's TorchTitan training feature. Meta is working with AMD, Arm, Dell, Intel, and NVIDIA to optimize performance across devices. The company has also released resources through its AI Developer Center for developers working with the model.

Meta Debuts New AI Model as CEO Shares 'AI for Everyone' Vision | PYMNTS.com

Mark Zuckerberg published a 6,510-word essay Monday (Aug. 10) outlining a future where everyone has "free or affordable access" to what he calls Meta's superintelligence tools, including personal artificial intelligence (AI) agents. "For everyone to be part of the future, everyone must have the ability to use superintelligence to improve their lives and shape the world," Zuckerberg wrote. "We will offer free versions that will be accessible to billions of people." Sharing the technology that widely, Zuckerberg argued, will "check and balance the power of institutions," governments and businesses included. "Most other labs are focused on building AI for companies, governments, or other institutions. So if those labs lead, then the balance of power will favor larger institutions over individuals," Zuckerberg added. The essay also calls for deeper collaboration between AI labs and the government, letting the government examine AI models earlier in the development process. "This way, the government gains a security capability without restricting or delaying individuals' access to personal superintelligence or causing an imbalance of power," Zuckerberg wrote. Meanwhile, Meta is instituting a new governance system which gives its board the authority to establish safety criteria for AI models and determine if new models meet those standards. "I do not think it is in my, Meta's, or the world's best interests for me or anyone else to be a sole decision maker on how superintelligence is deployed," he wrote, adding that there should be "an industrywide version of this process." Also Monday, Meta debuted Muse Glimmer, the next model from its Meta Superintelligence Labs, saying it was open sourcing the model weights. "Muse Glimmer is a 30-billion-parameter model optimized for always-on local agent workflows," the company said in its announcement. "It's small enough to run on a Mac or PC with a single consumer GPU, enabling use cases that range from local agents and function calling, to local coding, and LLM-as-a-judge evaluation." Meta last week introduced the beta version of a terminal coding agent called Muse Code, powered by Muse Spark 1.2, a coding-focused model update. As PYMNTS noted at the time, the company has been facing pressure to show it can monetize AI tools such as its Muse Spark model and to provide meaningful growth to justify its enormous capital expenditures on AI.

Mark Zuckerberg dispels doomsday fears, heralds a rosy AI dawn

In a blog post, Zuckerberg said, "I do not understand why anyone who believes that AI will eliminate most jobs and much of humanity's relevance would rush to build that future. The notion that AI is so dangerous that the only safe path is an extreme concentration of power seems inherently problematic." Contradicting the doomsday outlook on artificial intelligence (AI), Meta's Mark Zuckerberg championed the technology and said that superintelligence, when achieved, will take knowledge and invention forward, and is not something that will simply eliminate jobs and make humans irrelevant. In a blog post, Zuckerberg said, "I do not understand why anyone who believes that AI will eliminate most jobs and much of humanity's relevance would rush to build that future. The notion that AI is so dangerous that the only safe path is an extreme concentration of power seems inherently problematic." He explained that the biggest sprints humanity has made have come from individual achievement and not from powerful institutions. Invention, not automation, will be the greatest contribution of superintelligence, Zuckerberg wrote, adding that early AI could do routine work, but soon it will help discover new knowledge. Referring to rising AI costs, he wrote, "While the number of questions a person can ask in a day is limited, the number of valuable things superintelligence can invent to help achieve your goals is unlimited." What Meta is doing Zuckerberg went on to explain how the Facebook and Instagram parent is developing superintelligence in a way that improves lives and is accessible widely. For instance, developing a capable personal agent which understands the user and her goals. "Your agent will work 24/7 on your behalf to improve your relationships, health, career, finances, home management, hobbies, and more. It will free up time for the things you enjoy, and help you accomplish more than you could otherwise." Then there are tools to express your ideas, build businesses, etc. Zuckerberg said researchers at Meta are using AI to generate new crystal structures that are ideal for augmented reality glasses, and engineers are creating new apps in a fraction of the time it would have taken before. "Everyone will soon have invention superpowers," he said. Broader idea In a detailed note on the Meta website around this time last year, describing the company's approach as distinct from other technology firms, Zuckerberg had outlined plans to deliver superintelligence that will help individuals achieve personal goals, create new things, and enhance daily life. "As profound as the abundance produced by AI may one day be, an even more meaningful impact on our lives will likely come from everyone having a personal superintelligence that helps you achieve your goals, create what you want to see in the world, experience any adventure, be a better friend to those you care about, and grow to become the person you aspire to be," he had said.

Muse Glimmer: Complete Guide to Meta's Open Agentic AI Model

Meta has released Muse Glimmer, a new open-weight AI model built for agent tasks, tool use, code work, image input, and long tasks. Meta released the model in August 2026 under the Apache 2.0 license. The model has about 29.6 billion parameters and uses a dense Transformer design. A 1.8 billion parameter vision encoder gives Glimmer the ability to read images, screens, charts, and documents. The model also supports more than 100 languages and has a 131,072-token context window. Its knowledge cutoff sits at January 4, 2026. targets a different need from a normal chatbot. The model aims to act as a local AI agent that can plan tasks, call tools, write code, inspect visual data, recover from errors, and handle several steps in one task. Meta created Glimmer from Muse Spark through a distillation process. Muse Spark remains the larger model, while Glimmer brings part of that capability into a much smaller package. Meta also plans an open-weight release of Muse Spark, which could make the Muse family far more important for the open AI market. Glimmer shows some of its best results on agent tests. Meta reports a 75.5 score on MCP Atlas, 74.6 on DeepSearch QA, 23.5 on τ3-Banking, 47.6 on WildClawBench, and 43.3 on Gaia2. On several of these tests, Glimmer scores above Gemma 4 31B and Qwen 3.6 27B. Yet Qwen 3.6 27B remains ahead on SkillsBench and OSWorld-Verified, with scores of 46.6 and 75.6 against Glimmer's 44.3 and 65.9. Also Read - How to Self-Host AI Agents on a VPS: Running Ollama & OpenClaw Muse Glimmer stands out less as another 30B chatbot and more as a serious attempt to put agent AI on local hardware. A 29.6B dense model, 131K context, image input, tool use, Apache 2.0 licensing, sub-20 GB quantization, and DFlash speed gains create a rare combination. The benchmark record remains mixed, yet the local agent use case looks unusually strong. Meta's release also signals a wider return to open-weight AI, with a future open-weight Muse Spark release potentially carrying even greater impact.

Meta launches new AI in push for best open source model to beat China

Meta released a new artificial intelligence model Monday as CEO Mark Zuckerberg called on Washington to lower barriers for US open-source developers to help them beat Chinese rivals. The Meta chief warned that American AI companies face restrictions that give foreign competitors an advantage, arguing that US policy should aim to make American open-source models the best in the world. "It is also important that the US and its allies lead the open source AI ecosystem that will make up a large percent of global AI use," Zuckerberg wrote Monday in a sweeping manifesto titled "The Future is for Everyone: The Path to a Positive AI Future." "Foreign labs currently hold several advantages here since American labs have to comply with many additional restrictions on training data," he added. Chinese open-weight models have meanwhile been gaining traction among US businesses because they offer increasingly competitive performance at a fraction of the cost of some American rivals -- even as policymakers in Washington weigh new restrictions over national-security concerns. Models from Chinese firms including DeepSeek, Alibaba and Z.ai have emerged as formidable low-cost alternatives, with some US tech companies embracing Chinese models for their flexibility and cheaper computing costs. But Washington has moved to curb their use in government. DeepSeek is barred from US intelligence-community systems, while lawmakers have floated broader restrictions on Chinese AI models and officials are considering additional curbs on Chinese open-weight technology.

Meta launches new AI model as Zuckerberg champions open-weight push

Meta CEO Mark Zuckerberg called for lower U.S. barriers for open-source AI models to better compete with Chinese rivals as the social media giant released a new open-weight model on Monday and said it plans to launch more such models soon. The new model, Muse Glimmer, is much smaller than leading AI models from rivals and is instead designed for agentic tasks and can run on a Mac or PC with a single graphics card, aiming to tap demand for AI systems that run directly on people's devices. Open-weight models are typically cheaper than leading models from so-called frontier labs such as OpenAI and Anthropic. They also come with publicly accessible core components for easy customization, unlike closed models that companies keep fully under their control. Meta's launch comes as the social media giant seeks to strengthen its position after forming a costly, new superintelligence team last year to propel itself back into the high-stakes AI race. Zuckerberg's statement also marks the latest show of support for open-weight AI, which is gaining traction as businesses grow wary of ballooning AI bills and worry about recent cybersecurity incidents involving models from Anthropic, OpenAI and Meta. Hugging Face, the AI coding collaboration site that was hacked by a rogue OpenAI model, said last month that it used a Chinese open-weight model to defend against the attack because closed-source models have restrictions on use for cybersecurity work. Shares of Meta, which have fallen about 10% so far this year, were up 1% in premarket trading on Monday. POLICY RETHINK NEEDED TO PROPEL OPEN-WEIGHT Zuckerberg said in a statement that the U.S. needed to rethink policies if domestic firms were to lead in open-weight models. Chinese startups are leading the race for open-weight models, with Moonshot's Kimi K3, alongside Alibaba's Qwen3.8-Max and DeepSeek's V4-Flash, delivering performance that rivals top systems by U.S. AI labs. By contrast, the leading models of U.S. developers OpenAI, Anthropic and Alphabet's Google are closed source. "Foreign labs currently hold several advantages here since American labs have to comply with many additional restrictions on training data," Zuckerberg said, referring to open-source models. * S&P 500 market updates here "U.S. policy must reduce this additional friction if we want American open source models to lead over time," Zuckerberg said, adding that restricting access to foreign open-source models was not an effective solution. U.S. President Donald Trump's administration told AI developers earlier this month that it will not put open-weight AI models through voluntary safety tests, according to two sources familiar with the discussions. In his statement, Zuckerberg also advocated for AI model distillation, or using a powerful AI system to train a smaller model. He said that Meta would implement a governance structure to give its independent directors the power to approve the safety criteria for releasing models. (Reporting by Ananya Palyekar and Shubham Kalia in Bengaluru; Editing by Mrigank Dhaniwala and Devika Syamnath)

Meta shares rise on Muse Glimmer launch and a radical AI vision By Investing.com

Investing.com -- Meta's stock popped 2.4% this morning following the launch of Muse Glimmer, a highly efficient, 30-billion-parameter AI model capable of running on a single GPU. Positioned as a lightweight, distilled version of the more powerful Muse Spark 1.2, Glimmer is tailor-made for "agent-like" tasks -- think schedule management, rapid prototyping, and file organization. By releasing the model weights for free on Hugging Face, Meta is directly countering the permissive licensing strategies of Chinese rivals like DeepSeek and Alibaba, while carving out a distinct lane against US giants (Amazon, Alphabet, Microsoft) who are largely focused on enterprise and government AI. The launch coincides with a massive philosophical declaration from CEO Mark Zuckerberg, signaling a pivot from centralized AI toward democratized, "Personal Superintelligence." Deep Dive: Key Details Hidden in the Essay Zuckerberg's lengthy essay outlines several major corporate policy shifts, infrastructure plans, and regulatory proposals that are highly consequential for investors and the tech sector. Here are the most important details you need to know: 1. A Major Shift in AI Governance Meta is stepping away from sole founder-control over AI safety. Zuckerberg announced that Meta is empowering its independent board of directors to approve safety criteria for model releases and review compliance. He is actively urging other frontier AI labs to adopt similar industry-wide oversight to avoid a single CEO having absolute authority. 2. A New Compromise for Government & National Security To balance the rapid release of open-source models with national security, Zuckerberg proposed a new framework: rather than delaying public releases for government review, Meta will share intermediate training checkpoints and technical staff with the US government before a model is finished. This allows the government to harden critical infrastructure early without throttling consumer access. 3. "Meta Superintelligence Labs" is Resuming Open-Source Releases Zuckerberg confirmed that the newly established "Meta Superintelligence Labs" are fully operational and that Meta will "resume releasing some open source models soon." He also fiercely defended "distillation" (AI models learning from other models) as a necessary practice for US competitiveness. 4. A New Cloud Business with "Dynamic Auctions" While Meta plans to offer free AI access to billions, it is officially standing up a cloud infrastructure business for heavy users. To price this, Meta will implement a dynamic auction mechanism for compute power, designed to guarantee users the lowest possible price based on real-time capacity and collective demand. 5. WhatsApp-Style Encryption for AI Agents Addressing privacy and government surveillance, Meta is building a "fully private mode" for its personal AI agents. Similar to WhatsApp's end-to-end encryption, this ensures that not even Meta (or the government) can access the data, tasks, or interactions handled by the user's agent. 6. Massive "Community Compacts" for Data Centers To overcome the massive friction of building AI infrastructure in the US, Meta is launching aggressive local incentive programs: * The Future Is For Everyone Fund: Direct financial injections into local communities (e.g., funding $50,000 bonuses for teachers in Richland Parish, Louisiana, where a data center is being built). * America's Workforce Academy: Free training and guaranteed high-paying jobs for skilled tradespeople (electricians, carpenters) to build out the physical grid. * Water & Energy Pledges: Meta aims to be 200% water-positive in highly stressed areas by 2030, and claims it is building its own energy-generating infrastructure that sometimes supplies surplus low-cost energy back to the local grid. 7. The "Balance of Power" Theory Zuckerberg's core argument against the prevailing "AI doom" narrative is that safety doesn't come from restricting AI, but from heavily proliferating it. He argues that giving everyone a superintelligent lawyer, cybersecurity agent, or business advisor creates a natural system of checks and balances -- preventing any single corporation, government, or AI from achieving tyrannical dominance.

Meta launches new AI model as Zuckerberg champions open-weight push

The new model, Muse Glimmer, is much smaller than leading AI models from rivals and is instead designed for agentic tasks and can run on a Mac or PC with a single graphics card, aiming to tap demand for AI systems that run ⁠directly ⁠on people's devices. Meta CEO Mark Zuckerberg called for lower U.S. barriers for open-source AI models to better compete with Chinese rivals as the social media giant released a new open-weight model on Monday and said it plans to launch more such models soon. The new model, Muse Glimmer, is much smaller than leading AI models from rivals and is instead designed for agentic tasks and can run on a Mac or PC with a single graphics card, aiming to tap demand for AI systems that run ⁠directly ⁠on people's devices. Open-weight models are typically cheaper than leading models from so-called frontier labs such as OpenAI and Anthropic. They also come with publicly accessible core components for easy customization, unlike closed models that companies keep fully under their control. Meta's launch comes as the social media giant seeks to strengthen its position after forming a costly, new superintelligence team last year to propel itself back into the high-stakes AI race. Zuckerberg's statement also marks the latest show of support for open-weight AI, which is gaining traction as businesses grow wary of ballooning ⁠AI bills and worry about recent cybersecurity incidents involving models from Anthropic, OpenAI and Meta. Hugging Face, the AI coding collaboration site that was hacked by a rogue OpenAI model, said last month that ⁠it used a Chinese open-weight model to defend against the attack because closed-source models have restrictions on use for cybersecurity work. Shares of Meta, which have fallen about 10% so far this year, were up 1% in premarket trading on Monday. Policy rethink needed to propel open-weight Zuckerberg said in a statement that the U.S. needed to rethink policies if domestic firms were to lead in open-weight models. Chinese startups are leading the race for open-weight models, with Moonshot's Kimi K3, alongside Alibaba's Qwen3.8-Max and DeepSeek's V4-Flash, delivering performance that rivals top systems by U.S. AI labs. By contrast, the leading models of U.S. developers OpenAI, Anthropic and Alphabet's Google are closed source. "Foreign labs currently hold several advantages here since American labs have to comply with many additional restrictions on training data," Zuckerberg said, referring to open-source models. "U.S. ⁠policy must reduce this additional friction if we want American open source models to lead over time," Zuckerberg said, adding that restricting access to foreign open-source models was not an effective solution. U.S. President Donald Trump's administration told AI developers earlier this month that it will not put open-weight AI models through voluntary safety tests, according to two sources familiar with the discussions. In his statement, Zuckerberg also advocated for AI model distillation, or using a powerful AI system to train a smaller model. He said that Meta would implement a governance structure to give its independent directors the power to approve the safety criteria for releasing models.

Muse Glimmer: Meta Goes Bigger on Open-Weight AI

Meta introduced a new open-weight model as CEO Mark Zuckerberg pushes for a wider shift toward systems that can be downloaded, customized and operated directly on personal computers. The move comes as the social media company tries to strengthen its position in a fiercely competitive technology market while US and Chinese developers race to advance their own models. The newly released Muse Glimmer is considerably smaller than the most advanced models offered by companies such as OpenAI and Anthropic. It can run on a Mac or PC equipped with a single graphics card, potentially making it more accessible to developers and businesses that want to .

Muse Glimmer explained: Meta's Open AI model that works on a single GPU

One of the most practical open-source projects by Meta so far is being silently released. The new Muse Glimmer model by Meta Superintelligence Labs is an agentive artificial intelligence architecture with 30 billion parameters that can operate entirely on your machine with no use of clouds, APIs, or internet at all. The project is out now under an Apache 2.0 license, with weights hosted on Hugging Face. Also read: GPT-5.6-Cyber explained: OpenAI's cybersecurity AI with fewer safety refusals In short, Muse Glimmer breaks the trend in which most AI agents are located in clouds since they have a lot of computational power and cannot be run locally. At the same time, this architecture is so light to fit into consumer GPUs and even Macs with Apple Silicon, but still can perform such tasks as calling commands, coding, debugging, multiple step-by-step reasoning, handling failures and even perceive pictures and documents thanks to perception encoding capability. How Meta shrank it down In its fullest precision, such a model would consume more than 55GB of memory, which is far beyond the capacities of any gaming GPU. As a solution to that problem, the researchers used aggressive quantization to bring the model down to 4 bits, allowing it to fit into a budget of less than 20GB. This allows the remaining 4-8GB to be used for the KV cache, the encoder to understand the images and the speculative decoding "drafter" model, which can all run simultaneously on an RTX 4090, RTX 5090, or an advanced MacBook with 24 or 32GB of memory. Also read: Yahoo is building an email inbox you never have to open The second hack comes from the drafter model. It relies on a new DFlash approach and is capable of proposing blocks of text at once rather than tokens, while the main model checks its proposals. According to the researchers' estimates, this significantly increases generation speed - up to 3.1x for an RTX 5090, 1.8x for an M5 Max and 1.5x for an M4 Max, with no decrease in the output quality. Why this matters For India's PC builder and gamers, this is arguably more of a fascinating story than another chatbot launch. One single high-performing graphics card, the kind you have for playing games such as Valorant and Cyberpunk, can now become a local AI agent that will not depend on cloud computing and will keep your personal data safe from any server. It is an honest proposition for programmers and enthusiasts of artificial intelligence who do not want to be dependent on the cloud. According to Meta, Muse Glimmer is competitive with other similar models such as Gemma4-31B and Qwen3.6-27B in the area of agency and coding tasks, however, it is still better to wait for benchmarking from independent third-party sources before taking these comparisons for granted. Integration into llama.cpp, MLX and ExecuTorch is coming very soon, as well as serving through vLLM, SGLang, Ollama, LM Studio, and OpenRouter. Meta is cooperating with AMD, Intel, Arm, Dell and Nvidia for improving the model's performance on various hardware. Muse Glimmer can prove itself on the benchmark test, showing that powerful agents can live even without a data center and just a good GPU at your disposal.

Meta launches Muse Glimmer, a 30B AI model designed for local AI agents

Muse Glimmer supports coding, tool use, multimodal inputs and long-horizon tasks, with weights released under the Apache 2.0 licence. Meta has introduced Muse Glimmer, a 30-billion-parameter AI model designed to run AI agents locally on consumer hardware. Mark Zuckerberg's company is releasing the model weights under the Apache 2.0 licence allowing developers to download, modify and deploy the model for their own applications. It is aimed at tasks such as coding, function calling, tool use and multi-step agent workflows. Meta says it can run on a Mac or PC with a single consumer GPU, reducing the need for cloud-based infrastructure or a constant internet connection. Muse Glimmer is for agentic AI Meta says Muse Glimmer has been trained to handle longer workflows, including tool calls, reasoning, coding and failure recovery. The model can identify failed tool calls and attempt to correct them instead of stopping the task. It also supports the multimodal inputs allowing agents to work with text and images such as screenshots, charts and documents. The model has been trained using data from more than 100 languages. Also read: Amazon Great Freedom Sale 2026: Best smartphone deals for under Rs 50,000 Meta used outputs from its larger Muse Spark model during pre-training through logit distillation. The company then added longer-context and agent-focused training, followed by supervised fine-tuning, reinforcement learning and further distillation across reasoning, coding and agentic tasks. Quantisation brings the model below 20GB To run a 30B model at full precision would require more than 55GB of memory, according to Meta. The company has therefore used quantisation to reduce the model to under 20GB at roughly 4-bit precision. This allows it to operate within systems offering 24GB or 32GB of memory while leaving room for its working memory and image-processing components. Meta is also using speculative decoding to improve response speeds. A smaller companion model proposes groups of tokens, which Muse Glimmer then verifies. Meta says this increased decoding speed by up to 3.1 times on an Nvidia RTX 5090, 1.8 times on an M5 Max and 1.5 times on an M4 Max. Muse Glimmer is available through Hugging Face. Support for platforms and frameworks including Ollama, LM Studio, llama.cpp, MLX and ExecuTorch is expected to follow.