Meta Unveils Llama 3.2: A Groundbreaking Open-Source Multimodal AI Model

Inside Llama 3.2's Vision Architecture: Bridging Language and Image Understanding

Meta's Llama 3.2 has been developed to redefined how large language models (LLMs) interact with visual data. By introducing a groundbreaking architecture that seamlessly integrates image understanding with language processing, the Llama 3.2 vision models -- 11B and 90B parameters -- push the boundaries of multimodal AI. This evolution not only broadens the scope of what AI can achieve but also opens up new possibilities for applications in industries ranging from healthcare to finance, and beyond. In this overview, we will explore how Llama 3.2's vision architecture works, and how it bridges the gap between image reasoning and natural language understanding. Quick Links Llama 3.2 is a multimodal model designed to understand both visual data and natural language through a tightly integrated architecture. At its core, the Llama 3.2 vision models (available in 11B and 90B parameters) leverage a pre-trained image encoder to process visual inputs, which are then passed through the language model. What sets Llama 3.2 apart from its predecessors is its ability to seamlessly merge these two data types. While many AI models excel in either vision or language tasks, Llama 3.2 excels at both, using cross-attention layers that connect the image representations with the language model's pre-trained text data. This results in enhanced cross-modal reasoning, where the model can deeply understand and generate natural language that corresponds to complex visual data. These capabilities are particularly useful in tasks like document understanding -- analyzing charts, graphs, or even images in legal documents -- where both the textual and visual content need to be processed together to generate meaningful insights. The key innovation in Llama 3.2's vision architecture is the cross-attention mechanism, which allows the model to attend to both image and text data simultaneously. Here's how it functions: Image Encoder: The image input is processed through a pre-trained image encoder, which extracts relevant features from the image. The encoder translates the raw visual data into a set of image representations, which can be interpreted by the model. Cross-Attention Layers: These image representations are then passed into the cross-attention layers, which align the visual data with the text-based data. Cross-attention enables the model to understand how textual descriptions relate to visual elements, allowing for more complex reasoning tasks. Text Model Integration: After the image features are processed, they are passed into the language model where they interact with the textual data. This combined representation enables Llama 3.2 to generate text that is contextually grounded in the image or visual content. The power of cross-attention lies in its ability to contextualize visual data within the broader narrative of a document or question. This architecture can reason about objects, scenes, and spatial relationships in an image, and then describe them accurately in natural language or answer specific questions about the visual content. Llama 3.2's robust architecture paves the way for several practical applications across different industries: 1. Document-Level Understanding The 11B and 90B models excel at interpreting visual data in documents, such as financial reports or legal documents containing charts and graphs. Llama 3.2 can analyze and interpret these visual elements, offering insights and generating meaningful summaries that combine both the textual and visual aspects of the document. 2. Image Captioning In the domain of media and content generation, Llama 3.2 offers image captioning capabilities that allow it to describe scenes or images in natural language. For instance, an AI-powered photo app can automatically generate a caption that accurately describes the contents of a user's photo, from landscapes to complex indoor settings. 3. Visual Question Answering (VQA) Llama 3.2's ability to answer questions about an image is particularly useful in fields like education and customer service. Imagine asking a system questions about a geographical map or an anatomical chart, and having it respond with precise, well-reasoned answers based on the visual data. 4. Healthcare and Medical Imaging Medical professionals can use Llama 3.2's vision models for tasks like interpreting X-rays, MRI scans, or histology slides. The model can generate text-based insights about a medical image, assisting in diagnostic decision-making while integrating patient history or additional textual data. 5. Retail and E-commerce In e-commerce, Llama 3.2 can enable image search capabilities, where users submit a photo of a product, and the model finds relevant information, descriptions, or similar products. It can also be used to automatically generate product descriptions by analyzing product images. The training pipeline for Llama 3.2's vision models is a multi-stage process that builds upon the pre-trained language model by adding visual understanding. Here's an overview of the steps involved: 1. Pre-training on Large-Scale Data Llama 3.2 was initially trained on large-scale, noisy image-text pair data to ensure a broad understanding of both visual and textual elements. This stage allowed the model to develop an initial alignment between images and their corresponding text. 2. Fine-tuning with Domain-Specific Data The next stage involved fine-tuning on high-quality, domain-specific data. For instance, models trained for healthcare use cases would be fine-tuned on medical images and corresponding reports, optimizing the model's performance in that specific domain. 3. Alignment and Safety Mitigations In post-training, Llama 3.2 undergoes several rounds of alignment, including supervised fine-tuning, rejection sampling, and preference optimization to enhance safety and user alignment. Synthetic data generation is used during this phase to further refine the model's outputs in multimodal tasks. Llama 3.2's ability to bridge the gap between vision and language represents a significant leap forward in multimodal AI. As applications for this technology continue to expand, we can expect to see even more sophisticated systems capable of reasoning about images and generating highly contextualized responses in various fields. From healthcare to content creation and beyond, Llama 3.2 is set to unlock new possibilities for AI that truly understands and interacts with the world as we do. Here are a selection of other articles from our extensive library of content you may find of interest on the subject of Llama 3.2 :

Llama 3.2: Meta's Next Leap in Vision AI

The release of Meta's Llama 3.2 has marked a significant advancement in the landscape of generative AI, particularly in the field of vision AI models. Llama 3.2 offers a blend of text and vision capabilities, setting new benchmarks in image reasoning, visual grounding, and text generation for on-device use. This breakthrough makes AI more accessible to developers and enterprises, especially with the robust infrastructure Meta has developed to support these models. In this overview, we will dive deep into the key aspects of Llama 3.2, exploring its core features, architecture, and what sets it apart from its predecessors. Meta's Llama 3.2 represents a monumental step in advancing multimodal AI capabilities, including both vision and text processing. What stands out about Llama 3.2 is its architecture, which combines vision and language models, offering pre-trained and instruction-tuned variants that are adaptable to multiple environments. The 11B and 90B models focus on vision tasks, while the lightweight 1B and 3B models are optimized for text-based tasks on mobile and edge devices. Llama 3.2 enables the ability to process 128K tokens, an unprecedented length in on-device models, making it ideal for tasks such as extended summarization and rewriting. It is also designed to integrate into popular hardware ecosystems, including Qualcomm, MediaTek, and Arm processors, offering real-time AI processing without compromising on privacy or speed. The vision models in particular offer superior performance on image understanding tasks compared to closed alternatives like Claude 3 Haiku, making Llama 3.2 a new contender in AI image processing. One of the most exciting developments in Llama 3.2 is its vision capabilities. The 11B and 90B models are designed specifically for image reasoning tasks, offering developers the ability to integrate visual understanding into their applications. These models can perform complex tasks such as document-level understanding (e.g., interpreting charts and graphs), image captioning, and even pinpointing objects in images based on natural language descriptions. For example, Llama 3.2 can analyze sales graphs to answer questions about business performance or reason over maps to provide hiking trail information. These capabilities provide a seamless bridge between text and image data, enabling a wide range of applications, from business analytics to navigation. In addition to the vision models, Llama 3.2 introduces smaller, more efficient text-only models -- 1B and 3B. These models are highly optimized for on-device use cases, including summarization, tool usage, and multilingual text generation. By using pruning and distillation techniques, Meta has made it possible to compress larger models while retaining significant performance. These lightweight models bring a new level of privacy to applications, as they allow data to be processed entirely on the device without needing to be sent to the cloud. This is particularly relevant for sensitive tasks like summarizing messages, extracting action items, or scheduling follow-up meetings. The combination of on-device processing with powerful tool-calling abilities opens new possibilities for developers who want to build personalized, privacy-focused applications. To make it easier for developers to deploy and scale Llama models, Meta has introduced the Llama Stack Distribution. This collection of tools simplifies the deployment of Llama 3.2 models in various environments, from single-node on-premises systems to cloud-based infrastructures. The Llama Stack includes pre-configured APIs for inference, tool use, and retrieval-augmented generation (RAG), enabling developers to focus on building applications rather than managing infrastructure. It also supports integration with leading cloud platforms like AWS, Databricks, and Fireworks, as well as on-device solutions via PyTorch ExecuTorch. By offering a standardized interface and client code in multiple programming languages, Llama Stack ensures that developers can easily transition between different deployment environments. As part of its commitment to responsible AI development, Meta has also introduced new safety features in Llama 3.2. The Llama Guard 3 11B Vision model includes safeguards that filter text and image inputs to ensure they comply with safety guidelines. Additionally, Llama Guard 3 1B has been pruned and quantized to make it more efficient for on-device deployment, drastically reducing its size from 2,858 MB to just 438 MB. These safeguards are critical for ensuring that AI applications built on Llama 3.2 adhere to best practices in privacy, security, and responsible innovation. Llama 3.2 provides developers with a robust and versatile platform for building AI applications. Whether it's creating agentic applications with tool-calling abilities, building privacy-focused on-device solutions, or scaling up cloud-based AI models, Llama 3.2's modular architecture supports a wide range of use cases. With its lightweight models optimized for mobile and edge devices, and its powerful vision models capable of complex image reasoning, Llama 3.2 will likely become a cornerstone for next-generation AI development. Additionally, Meta's strong partnerships with leading tech companies like AWS, Qualcomm, and Google Cloud ensure that developers have the support and infrastructure they need to implement these models at scale. Llama 3.2's focus on openness and modifiability offers a transparent, community-driven approach to AI, empowering more innovators to experiment and develop cutting-edge solutions.

New Meta Llama 3.2 Open Source Multimodal LLM Launches

Meta AI has unveiled the Llama 3.2 model series, a significant milestone in the development of open-source multimodal large language models (LLMs). This series encompasses both vision and text-only models, each carefully optimized to cater to a wide array of use cases and devices. Llama 3.2 comes in two primary variants: This versatility allows users to select the model that aligns perfectly with their specific requirements, ensuring optimal performance and efficiency across various applications. Llama 3.2 has demonstrated remarkable performance, surpassing leading models such as CLA 3 Haiku and GPT 4 Omni mini in numerous benchmarks. Its exceptional capabilities shine in tasks like image captioning, visual question answering (VQA), and image-text retrieval. These benchmarks underscore the model's superior proficiency in both vision and text tasks, establishing it as a versatile and powerful tool for a wide range of applications. Moreover, Llama 3.2 is designed with speed and accuracy in mind, supporting up to 128k tokens. This enables the model to tackle extensive tasks, such as summarization and instruction following, with unparalleled efficiency. The model's optimization for various processors ensures seamless compatibility and optimal performance across different hardware platforms, making it a practical choice for real-world deployments. Llama 3.2 introduces a groundbreaking architecture that seamlessly integrates a pre-trained image encoder with a language model using cross-attention layers. This innovative design significantly enhances the model's ability to process and understand multimodal data, unlocking new possibilities for complex tasks involving both vision and language. The training pipeline of Llama 3.2 incorporates several key elements, including: These techniques collectively contribute to the model's exceptional performance and adaptability, allowing it to excel in a wide range of applications and domains. Here are a selection of other articles from our extensive library of content you may find of interest on the subject of Meta Llama AI models : Recognizing the growing demand for on-device AI capabilities, Llama 3.2 offers lightweight models created through advanced pruning and distillation techniques. These models maintain strong performance while being more efficient and compact, making them ideal for deployment on edge and mobile devices. This ensures that users can harness the power of innovative AI technologies even in resource-constrained environments, opening up new possibilities for innovative applications. Llama 3.2 models are readily available on popular platforms like Hugging Face and Together AI, ensuring easy access for developers and researchers. Additionally, users can install the models locally using platforms such as LM Studio, providing flexibility and convenience in deployment. The practical applications of Llama 3.2 are vast and diverse. One compelling example is its use in analyzing and categorizing data from receipts, showcasing the model's proficiency in both image understanding and textual prompts. This highlights the model's potential to transform various industries, from finance and retail to healthcare and beyond. The release of Llama 3.2 represents a significant leap forward for the open-source community. By providing a powerful and versatile multimodal LLM, Meta AI is helping to bridge the gap between open-source and closed-source models. This advancement fosters greater collaboration, knowledge sharing, and innovation within the community, driving the development of groundbreaking AI technologies that have the potential to transform industries and improve lives. As researchers, developers, and businesses explore the capabilities of Llama 3.2, we can expect to witness a surge in innovative applications and solutions that use the power of multimodal AI. With its exceptional performance, flexibility, and accessibility, Llama 3.2 is poised to become a fantastic option for the next generation of intelligent systems, propelling us towards a future where AI seamlessly integrates with and enhances various aspects of our lives.

Meta has officially released Llama 3.2

Meta has announced the production release of Llama 3.2, an unprecedented collection of free and open-source artificial intelligence models aimed at shaping the future of machine intelligence with flexibility and efficiency. Since businesses are on the lookout for apocalyptic AI solutions that can work on the hardware most common or on those that are popular to develop solutions for large businesses as well as independents, Llama 3.2 gives new models. An emphasis on the edge and mobility is something that is quite evident at Meta. As for the new features of this version, the developers have added small- and medium-sized vision LLM: 11B and 90B, and so also introduced pure text alternatives, 1B and 3B. Particularly, the new models introduced here are aligned for the operation of edge devices, thus making the AI technology available to more clients. The lightweight text-only models, especially those without any visual data, are designed for simpler tasks such as summarization and instruction following due to the low computation power. Due to central data processing on mobile devices, with local execution, none of the data is uploaded to the cloud, as Meta states, "Running locally on mobile devices ensures that the data remains on the device, enhancing user privacy by avoiding cloud-based processing," This capability is especially useful for applications that process sensitive data, as it enables the application to perform important tasks while maintaining the confidentiality of the data. For example, users can reply to personal messages while summarizing them, or get to-do-list items from meetings without relaying messages to external servers. The most significant change in Llama 3.2 is various architectural improvements. The new models use an adapter-based architecture that can combine image encoders with pre-trained text models without modification. This integration leads to improvements in the ability to reason in both text and image areas and greatly expands the range of applications for these models. The resulting pre-trained models went through stringent fine-tuning exercises that entailed the utilization of huge noisy image-text pair data. There is one important addition to the token context length, and it increased to a very impressive 128K for the lightweight 1B and 3B models. It facilitates wider data travelling which is particularly valuable for long documents and elaborate thinking. This capability to accommodate such large input sizes places Llama 3.2 at an advantage with respect to competitors in the dynamic AI market dominated by OpenAI's GPT models. Llama 3.2's models have demonstrated exceptional performance metrics, further solidifying their competitive edge in the market. The 1B model achieved a score of 49.3 on the MMLU benchmark, while the 3B model scored 63.4. On the vision side, the 11B and 90B models showcased their capabilities with scores of 50.7 and 60.3, respectively, in visual reasoning tasks. These metrics indicate that the Llama 3.2 models not only meet but often exceed the performance of similar offerings from other companies, such as Claude 3 Haiku and GPT4o-mini. The integration of UnslothAI technology also adds to the efficiency of these models, enabling twice as fast fine-tuning and inference speeds while reducing VRAM usage by 70%. This enhancement is crucial for developers looking to implement real-time AI solutions without facing hardware limitations. One of the key factors that define Llama 3.2's readiness to be brought into the market is its well-developed ecosystem. Partnerships with other mobile industry leaders like Qualcomm, MediaTek, and AWS make it possible for developers to implement these models across different settings, cloud environments, and local devices. The Llama Stack distributions such as Llama Stack for on-device installations and Llama Stack for single-node installation offer solutions that developers can take advantage of and build these models into their projects without added complications. The latest version of the open-source AI model, Llama 3.2, is now available on the Meta Llama website, offering enhanced capabilities for customization, fine-tuning, and deployment across various platforms. Developers can choose from four model sizes: 1B, 3B, 11B, and 90B, or continue utilizing the earlier Llama 3.1. Meta is not just releasing these models into the wild; they are keen on ensuring developers have everything they need to leverage Llama 3.2 effectively. This commitment includes sharing valuable tools and resources to help developers build responsibly. By continuously updating their best practices and engaging with the open-source community, Meta hopes to inspire innovation while promoting ethical AI usage. "We're excited to continue the conversations we're having with our partners and the open-source community, and as always, we can't wait to see what the community builds using Llama 3.2 and Llama Stack," Meta stated. This collaborative approach not only enhances the capabilities of Llama 3.2 but also encourages a vibrant ecosystem. Whether for lightweight edge solutions or more complex multimodal tasks, Meta hopes that the new models will provide the flexibility needed to meet diverse user demands.

Meta's Llama 3.2 launches with vision to rival OpenAI, Anthropic

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Today at Meta Connect, the company rolled out Llama 3.2, its first major vision models that understand both images and text. Llama 3.2 includes small and medium-sized models (at 11B and 90B parameters), as well as more lightweight text-only models (1B and 3B parameters) that fit onto select mobile and edge devices. Like its predecessor, Llama 3.2 has a 128,000 token context length, meaning users can input lots of text (on the scale of hundreds of pages of a textbook). Higher parameters also typically indicate that models will be more accurate and can handle more complex tasks. Meta is today for the first time sharing official Llama stack distributions so that developers can work with the models in a variety of environments, including on-prem, on-device, cloud and single-node. Now, the two largest Llama 3.2 models (11B and 90B) support image use cases, and have the ability to understand charts and graphs, caption images and pinpoint objects from natural language descriptions. For example, a user could ask in what month their company saw the best sales, and the model will reason an answer based on available graphs. The larger models can also extract details from images to create captions. The lightweight models, meanwhile, can help developers build personalized agentic apps in a private setting -- such as summarizing recent messages or sending calendar invites for follow-up meetings. Meta says that Llama 3.2 is competitive with Anthropic's Claude 3 Haiku and OpenAI's GPT4o-mini on image recognition and other visual understanding tasks. Meanwhile, it outperforms Gemma and Phi 3.5-mini in areas such as instruction following, summarization, tool use and prompt rewriting. Also today, Meta is expanding its business AI so that enterprises can use click-to-message ads on WhatsApp and Messenger and build out agents that answer common questions, discuss product details and finalize purchases. The company claims that more than 1 million advertisers use its generative AI tools, and that 15 million ads were created with them in the last month. On average, ad campaigns using Meta gen AI saw 11% higher click-through rate and 7.6% higher conversion rate compared to those that didn't use gen AI, Meta reports. Finally, for consumers, Meta AI now has "a voice" -- or more like several. The new Llama 3.2 supports new multimodal features in Meta AI, most notably, its capability to talk back in celebrity voices including Dame Judi Dench, John Cena, Keegan Michael Key, Kristen Bell and Awkwafina. The model will respond to voice or text commands in celebrity voices across WhatsApp, Messenger, Facebook and Instagram. Meta AI will also be able to reply to photos shared in chat and add, remove or change images and add new backgrounds. Meta says it is also experimenting with new translation, video dubbing and lip syncing tools for Meta AI.

Meta drops multimodal Llama 3.2 -- here's why it's such a big deal

Meta has just dropped a new version of its Llama family of large language models. The updated Llama 3.2 introduces multimodality, enabling it to understand images in addition to text. It also brings two new 'tiny' models into the family. Llama is significant -- not necessarily because it's more powerful than models from OpenAI or Google, although it does give them a run for their money -- but because it's open source and accessible to nearly anyone with relative ease. The update introduces four different model sizes. The 1 billion parameter model runs comfortably on an M3 MacBook Air with 8GB of RAM, while the 3 billion model also works but just barely. These are both text only but can be run on a wider range of devices and offline. The real breakthrough, though, is with the 11b and 90b parameter versions of Llama 3.2. These are the first true multimodal Llama models, optimized for hardware and privacy and far more efficient than their 3.1 predecessors. The 11b model could even run on a good gaming laptop. Llama's wide availability, state-of-the-art capability, and adaptability set it apart. It powers Meta's AI chatbot across Instagram, WhatsApp, Facebook, Ray-Ban smart glasses, and Quest headsets, but it is also accessible on public cloud services, so users can download and run it locally or even integrate it into third-party products. Groq, the ultra-fast cloud inference service, is one example of why having an open-source model is a powerful choice. I built a simple tool to summarize an AI research paper using Llama 3.1 70b running on Groq - it completed the summary faster than I could read the title. Some open-source libraries let you create a ChatGPT-like interface on your Mac powered by Llama 3.2 or other models, including the image analysis capabilities if you have enough RAM. However, I took it a step further and built my own Python chatbot that queries the Ollama API, enabling me to run these models directly in the terminal. One of the significant reasons Llama 3.2 is such a big deal is its potential to transform how AI interacts with its environment, especially in areas like gaming and augmented reality. The multimodal capabilities mean Llama 3.2 can both "see" and "understand" visual inputs alongside text, opening up possibilities like dynamic, AI-powered NPCs in video games. Imagine a game where NPCs aren't just following pre-scripted dialogue but can perceive the game world in real-time, responding intelligently to player actions and the environment. For example, a guard NPC could "see" the player holding a specific weapon and comment on it, or an AI companion might react to a change in the game's surroundings, such as the sudden appearance of a threat, in a nuanced and conversational way. Beyond gaming, this technology can be used in smart devices like the Ray-Ban smart glasses and Quest headsets. Imagine pointing your glasses at a building and asking the AI for architectural history or details about a restaurant's menu just by looking at it. These use cases are exciting because Llama's open-source nature means developers can customize and scale these models for countless innovative applications, from education to healthcare, where AI could assist visually impaired users by describing their environment. Outside of using the models as built by Meta, being open-source means companies, organizations, and even governments can create their own customized and fine-tuned versions of the models. This is already happening in India to save near-extinct languages. Llama 3.2 11b and 90b are competitive with smaller models from Anthropic, such as Claude 3 Haiku and OpenAI, including GPT-4o-mini, when recognizing an image and similar visual tasks. The 3B version is competitive with similar-sized models from Microsoft and Google, including Gemini and Phi 3.5-mini across 150 benchmarks. While not a direct benchmark, my own tests of having the 1b model analyze my writing and offer suggested improvements are roughly equal to the performance of Apple Intelligence writing tools, just without the handy context menu access. The two vision models, 11b and 90b, can perform many of the same functions I've seen from ChatGPT and Gemini. For example, you can give it a photograph of your garden, and it can offer suggested improvements or even a planting schedule. As I've said before, the performance, while good, isn't the most significant selling point for Llama 3.2; it is in how accessible it is and customizable for a range of use cases.

Meta Releases Llama 3.2 Models with Vision Capability For the First Time

You can start using Llama 3.2 11B and 90B vision models through the Meta AI chatbot on the web, WhatsApp, Facebook, Instagram, and Messenger. At the Meta Connect 2024 event, Mark Zuckerberg announced the new Llama 3.2 family of models to take on OpenAI's o1 and o1 mini models. Moreover, for the first time, the Llama 3.2 models come with multimodal image support. First of all, Llama 3.2 has two smaller models, which include Llama 3.2 1B and 3B for on-device tasks. Meta says these small models are optimized to work on mobile devices and laptops. Llama 3.2 1B and 3B models are best suited for on-device summarization, instruction following, rewriting, and even function calling to create an action intent locally. Meta also claims that its latest Llama models outperform Google's Gemma 2 2.6B and Microsoft's Phi-3.5-mini. Basically, developers can deploy these models on Qualcomm and MediaTek platforms to power many AI use cases. Meta further says Llama 3.2 1B and 3B models are pruned and distilled from the larger Llama 3.1 8B and 70B models. Now coming to the exciting vision models, they come in larger sizes -- Llama 3.2 11B and Llama 3.2 90B. They replace the older text-only Llama 3.1 8B and 70B models. Meta goes on to say that Llama 3.2 11B and 90B models rival closed models like Anthropic's Claude 3 Haiku and OpenAI's GPT-4o mini in visual reasoning. These new Llama 3.2 11B and 90B vision models will be available through the Meta AI chatbot on the web, WhatsApp, Instagram, Facebook, and Messenger. Since these are vision models, you can upload images and ask questions about them. For example -- you can upload an image of a recipe, and it can analyze and give you instructions on how to make it. you can have Meta AI capture your face and reimagine yourself in tons of different scenarios and portraits. The vision models also come in handy while understanding charts and graphs. On social media apps like Instagram and WhatsApp, the vision models can also generate captions for you. Overall, Meta has released multimodal models for the first time under an open-source license. It is going to be pretty exciting to test the vision models against the competition.

Llama 3.2: Revolutionizing edge AI and vision with open, customizable models

We've been excited by the impact the Llama 3.1 herd of models have made in the two months since we announced them, including the 405B -- the first open frontier-level AI model. While these models are incredibly powerful, we recognize that building with them requires significant compute resources and expertise. We've also heard from developers who don't have access to these resources and still want the opportunity to build with Llama. As Meta Founder and CEO Mark Zuckerberg shared today at Connect, they won't have to wait any longer. Today, we're releasing Llama 3.2, which includes small and medium-sized vision LLMs (11B and 90B) and lightweight, text-only models (1B and 3B) that fit onto select edge and mobile devices. It's only been a year and a half since we first announced Llama, and we've made incredible progress in such a short amount of time. This year, Llama has achieved 10x growth and become the standard for responsible innovation. Llama also continues to lead on openness, modifiability, and cost efficiency, and it's competitive with closed models -- even leading in some areas. We believe that openness drives innovation and is the right path forward, which is why we continue to share our research and collaborate with our partners and the developer community. We're making Llama 3.2 models available for download on llama.com and Hugging Face, as well as available for immediate development on our broad ecosystem of partner platforms. Partners are an important part of this work, and we've worked with over 25 companies, including AMD, AWS, Databricks, Dell, Google Cloud, Groq, IBM, Intel, Microsoft Azure, NVIDIA, Oracle Cloud, and Snowflake, to enable services on day one. For the Llama 3.2 release, we're also working with on-device partners Arm, MediaTek, and Qualcomm to offer a broad range of services at launch. Starting today, we're also making Llama Stack available to the community. More details on the latest release, including information on the multimodal availability in Europe, can be found in our acceptable use policy.

Meta's Llama AI models get multimodal

Benjamin Franklin once wrote that nothing is certain except death and taxes. Let me amend that phrase to reflect the current AI goldrush: Nothing is certain except death, taxes, and new AI models, with the last of those three arriving at an ever-accelerating pace. Meta's multilingual Llama family of models has reached version 3.2, with the bump from 3.1 signifying that several Llama models are now multimodal. Llama 3.2 11B -- a compact model -- and 90B, which is a larger, more capable model, can interpret charts and graphs, caption images, and pinpoint objects in pictures given a simple description. Given a map of a park, for example, Llama 3.2 11B and 90B might be able to answer questions like, "When will the terrain become steeper?" and "What's the distance of this path?" Or, provided a graph showing a company's revenue over the course of a year, the models could quickly spotlight the best-performing months of the bunch. For developers who wish to use the models strictly for text applications, Meta says that Llama 3.2 11B and 90B were designed to be "drop-in" replacements for 3.1. 11B, and 90B can be deployed with or without a new safety tool, Llama Guard Vision, that's designed to detect potentially harmful (i.e. biased or toxic) text and images fed to or generated by the models. In most of the world, the multimodal Llama models can be downloaded from and used across a wide number of cloud platforms, including Hugging Face, Microsoft Azure, Google Cloud, and AWS. Meta's also hosting them on the official Llama site, Llama.com, and using them to power its AI assistant, Meta AI, across WhatsApp, Instagram, and Facebook. But Llama 3.2 11B and 90B can't be accessed in Europe. As a result, several Meta AI features available elsewhere, like image analysis, are disabled for European users. Meta once again blamed the "unpredictable" nature of the bloc's regulatory environment. Meta has expressed concerns about -- and spurned a voluntary safety pledge related to -- the AI Act, the EU law that establishes a legal and regulatory framework for AI. Among other requirements, the AI Act mandates that companies developing AI in the EU commit to charting whether their models are likely to be deployed in "high-risk" situations, like policing. Meta fears that the "open" nature of its models, which give it little insight into how the models are being used, could make it challenging to adhere to the AI Act's rules. Also at issue for Meta are provisions in the GDPR, the EU's broad privacy law, pertaining to AI training. Meta trains models on the public data of Instagram and Facebook users who haven't opted out -- data that in Europe is subject to GDPR guarantees. EU regulators earlier this year requested that Meta halt training on European user data while they assessed the company's GDPR compliance. Meta relented, while at the same time endorsing an open letter calling for "a modern interpretation" of GDPR that doesn't "reject progress." Earlier this month, Meta said that it would resume training on U.K. user data after "[incorporating] regulatory feedback" into a revised opt-out process. But the company has yet to share an update on its training throughout the rest of the bloc. More compact models Other new Llama models -- models that weren't trained on European user data -- are launching in Europe (and globally) Wednesday. Llama 3.2 1B and 3B, two lightweight, text-only models designed to run on smartphones and other edge devices, can be applied to tasks such as summarizing and rewriting paragraphs (e.g. in an email). Optimized for Arm hardware from Qualcomm and MediaTek, 1B and 3B can also tap tools such as calendar apps with a bit of configuration, Meta says, allowing them to take actions autonomously. There isn't a follow-up, multimodal or not, to the flagship Llama 3.1 405B model released in August. Given 405B's massive size -- it took months to train -- it's likely a matter of constrained compute resources. We've asked Meta if there are other factors at play and will update this story if we hear back. Meta's new Llama Stack, a suit of Llama-focused dev tools, can be used to fine-tune all the Llama 3.2 models: 1B, 3B, 11B, and 90B. Regardless of how they're customized, the models can process up to around 100,000 words at once, Meta says. A play for mindshare Meta CEO Mark Zuckerberg often talks about ensuring all people have access to the "benefits and opportunities" of AI. Implicit in this rhetoric, however, is a desire that these tools and models be of Meta's making. Spending on models that it can then commoditize forces the competition (e.g. OpenAI, Anthropic) to lower prices, spreads Meta's version of AI broadly, and lets Meta incorporate improvements from the open source community. Meta claims that its Llama models have been downloaded over 350 million times and are in use by large enterprises including Zoom, AT&T, and Goldman Sachs. For many of these developers and companies, it's immaterial that the Llama models aren't "open" in the strictest sense. Meta's license constrains how certain devs can use them; platforms with over 700 million monthly users must request a special license from Meta that the company will grant on its discretion. Granted, there aren't many platforms of that size without their own in-house models. But Meta isn't being especially transparent about the process. When I asked the company this month whether it had approved a discretionary Llama license for a platform yet, a spokesperson told me that Meta "didn't have anything to share on the topic." Make no mistake, Meta's playing for keeps. It's spending millions lobbying regulators to come around to its preferred flavor of "open" AI, and it's ploughing billions into servers, datacenters, and network infrastructure to train future models. None of the Llama 3.2 models solves the overriding problems with today's AI, like its tendency to make things up and regurgitate problematic training data (e.g. copyrighted ebooks that might've been used without permission, the subject of a class action lawsuit against Meta). But, as I've written before, they do advance one of Meta's key goals: becoming synonymous with AI, and in particular generative AI.

How Llama 3.2 is Transforming Edge Computing and On-Device AI

Meta's latest release of the Llama 3.2 model marks a significant advancement in AI, particularly in edge computing and on-device AI. Llama 3.2 brings powerful generative AI capabilities to mobile devices and edge systems by introducing highly optimized, lightweight models that can run without relying on cloud infrastructure. With the 1B and 3B text-only models, Meta ensures that users and developers can take advantage of AI in real-time, on-device environments while maintaining strong privacy and low latency. Quick Links Llama 3.2 is transforming the AI landscape by bringing cutting-edge capabilities to mobile and edge devices. While most AI models require heavy cloud-based infrastructure for processing, Llama 3.2's lightweight models (1B and 3B parameters) are specifically designed for deployment on edge devices -- hardware that operates closer to the data source, such as smartphones, IoT devices, and embedded systems. The edge-centric nature of Llama 3.2 means AI models can now run on devices themselves, instead of requiring continuous cloud connectivity. This shift enables: Real-time AI processing: Tasks like text generation, summarization, and rewriting can now be handled instantaneously on the device. Enhanced privacy: By processing data locally, sensitive user information remains on the device, reducing exposure to cloud-based vulnerabilities. Llama 3.2 has been optimized for Qualcomm, MediaTek, and Arm processors, making it versatile and efficient in various on-device environments. The ability to support a 128K token context length on edge devices is also groundbreaking, enabling complex tasks like summarization of long documents and instruction following to happen directly on the device. On-device AI has multiple advantages over traditional cloud-based AI, especially when powered by a model like Llama 3.2. Here are the primary benefits: 1. Low Latency and Instantaneous Responses One of the most significant advantages of on-device AI is its speed. Since the model runs locally, there's no need for data to travel back and forth between the device and the cloud. This results in faster response times, especially for tasks that require real-time interactions, such as voice assistants, augmented reality (AR) applications, and real-time analytics. 2. Improved Privacy and Data Security With on-device AI, data stays on the device, ensuring a higher level of privacy. Users don't have to worry about sensitive information being transmitted over the internet to cloud servers. This makes Llama 3.2 particularly valuable for applications involving personal data, such as messaging apps, email summarization, or healthcare applications. 3. Reduced Bandwidth and Operational Costs Running AI models on the cloud can lead to higher costs due to data transfer fees and the need for continuous connectivity. Edge computing eliminates these costs by allowing devices to handle their own processing, reducing reliance on bandwidth and saving operational costs. 4. Offline Functionality By shifting AI capabilities to edge devices, users can still benefit from AI-powered features even when they don't have internet access. This is particularly important for regions with poor connectivity or for applications that need to run in offline environments, such as remote industrial setups or autonomous vehicles. Llama 3.2's focus on lightweight, efficient models makes it perfect for a wide array of edge computing and on-device AI applications. Here are some notable examples: 1. Personalized Digital Assistants With Llama 3.2's ability to perform advanced tasks like text summarization and instruction following, developers can create highly personalized digital assistants that operate entirely on the user's device. These assistants can summarize emails, schedule meetings, and even generate custom responses without sending data to the cloud, making them faster and more private. 2. Smart IoT Devices The Internet of Things (IoT) continues to grow, and Llama 3.2's small-footprint models are ideal for deployment on smart devices like home assistants, wearables, and industrial sensors. These devices can now leverage real-time AI for tasks such as language understanding, predictive maintenance, or intelligent automation in factory settings, all while maintaining low power consumption. 3. Real-Time Analytics in Retail and Healthcare Retail systems can leverage Llama 3.2 for real-time customer analytics at the edge, such as analyzing consumer behavior or adjusting in-store promotions dynamically based on real-time data. Similarly, in healthcare, on-device AI can assist with diagnostics or real-time monitoring in remote health scenarios, without needing a constant internet connection. 4. Autonomous Vehicles Llama 3.2 can be integrated into autonomous systems where real-time decision-making is critical. Autonomous cars, drones, and robotics can process large amounts of data on-device, enabling faster reactions and enhanced situational awareness without depending on cloud-based processing, which is prone to delays. Llama 3.2 brings several technical innovations to the table, making it a leader in edge AI solutions. The most notable include: 1. Efficient Model Pruning and Distillation Meta used advanced pruning and distillation techniques to reduce the size of the models without compromising performance. The 1B and 3B parameter models are powerful enough to perform complex tasks while being small enough to run on mobile devices or edge servers. 2. 128K Token Context Length Llama 3.2 supports a context length of up to 128K tokens, even on edge devices, which is unprecedented for lightweight models. This allows developers to work with much larger contexts in summarization and document processing tasks, opening up possibilities for advanced use cases in edge environments. 3. Compatibility with Leading Hardware Platforms Llama 3.2 is optimized for Qualcomm, MediaTek, and Arm processors, ensuring that it can run efficiently on the most popular mobile and edge hardware. This broad compatibility ensures widespread adoption across different industries and verticals. By leveraging these technical advancements, Llama 3.2 is driving the next phase of on-device AI, transforming what's possible with edge computing. Meta's Llama 3.2 is not just another generative AI model -- it's a foundational technology for edge computing. By allowing powerful AI applications to run directly on devices, Llama 3.2 opens the door to faster, more private, and more versatile on-device AI, making it a game-changer in industries ranging from IoT to retail, healthcare, and beyond. Here are a selection of other articles from our extensive library of content you may find of interest on the subject of Edge Computing :

Meta rolls out its first major vision models to rival Anthropic, OpenAI

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Today at Meta Connect, the company rolled out Llama 3.2, its first major vision models that understand both images and text. Llama 3.2 includes small and medium-sized models (at 11B and 90B parameters), as well as more lightweight text-only models (1B and 3B parameters) that fit onto select mobile and edge devices. Like its predecessor, Llama 3.2 has a 128,000 token context length, meaning users can input lots of text (on the scale of hundreds of pages of a textbook). Higher parameters also typically indicate that models will be more accurate and can handle more complex tasks. Meta is today for the first time sharing official Llama stack distributions so that developers can work with the models in a variety of environments, including on-prem, on-device, cloud and single-node. Now, the two largest Llama 3.2 models (11B and 90B) support image use cases, and have the ability to understand charts and graphs, caption images and pinpoint objects from natural language descriptions. For example, a user could ask in what month their company saw the best sales, and the model will reason an answer based on available graphs. The larger models can also extract details from images to create captions. The lightweight models, meanwhile, can help developers build personalized agentic apps in a private setting -- such as summarizing recent messages or sending calendar invites for follow-up meetings. Meta says that Llama 3.2 is competitive with Anthropic's Claude 3 Haiku and OpenAI's GPT4o-mini on image recognition and other visual understanding tasks. Meanwhile, it outperforms Gemma and Phi 3.5-mini in areas such as instruction following, summarization, tool use and prompt rewriting. Also today, Meta is expanding its business AI so that enterprises can use click-to-message ads on WhatsApp and Messenger and build out agents that answer common questions, discuss product details and finalize purchases. The company claims that more than 1 million advertisers use its generative AI tools, and that 15 million ads were created with them in the last month. On average, ad campaigns using Meta gen AI saw 11% higher click-through rate and 7.6% higher conversion rate compared to those that didn't use gen AI, Meta reports. Finally, for consumers, Meta AI now has "a voice" -- or more like several. The new Llama 3.2 supports new multimodal features in Meta AI, most notably, its capability to talk back in celebrity voices including Dame Judi Dench, John Cena, Keegan Michael Key, Kristen Bell and Awkwafina. The model will respond to voice or text commands in celebrity voices across WhatsApp, Messenger, Facebook and Instagram. Meta AI will also be able to reply to photos shared in chat and add, remove or change images and add new backgrounds. Meta says it is also experimenting with new translation, video dubbing and lip syncing tools for Meta AI.

Meta releases its first open AI model that can process images

Just two months after releasing its last big AI model, Meta is back with a major update: its first open-source model capable of processing both images and text. The new model, Llama 3.2, could allow developers to create more advanced AI applications, like augmented reality apps that provide real-time understanding of video, visual search engines that sort images based on content, or document analysis that summarizes long chunks of text for you. Meta says it's going to be easy for developers to get the new model up and running. Developers will have to do little except add this "new multimodality and be able to show Llama images and have it communicate," Ahmad Al-Dahle, vice president of generative AI at Meta, told The Verge. Other AI developers, including OpenAI and Google, already launched multimodal models last year, so Meta is playing catch-up here. The addition of vision support will also play a key role as Meta continues to build out AI capabilities on hardware like its Ray-Ban Meta glasses. Llama 3.2 includes two vision models (with 11 billion parameters and 90 billion parameters) and two lightweight text-only models (with 1 billion parameters and 3 billion parameters). The smaller models are designed to work on Qualcomm, MediaTek, and other Arm hardware, with Meta clearly hoping to see them put to use on mobile. There's still a place for the (slightly) older Llama 3.1, though: that model, released in July, included a version with 405 billion parameters, which will theoretically be more capable when it comes to generating text.

AWS Makes Meta's Llama 3.2 LLMs Available to Customers | PYMNTS.com

This availability will offer AWS customers more options for building, deploying and scaling generative artificial intelligence (AI) applications, Amazon said in a Wednesday (Sept. 25) update. "The Llama 3.2 collection builds on the success of previous Llama models to offer new, updated and highly differentiated models, including small and medium-sized vision LLMs that support image reasoning and lightweight, text-only models optimized for on-device use cases," the update said. "The new models are designed to be more accessible and efficient, with a focus on responsible innovation and safety." The collection includes Llama 3.2 11B Vision and Llama 3.2 90B Vision, which are Meta's first multimodal vision models; Llama 3.2 1B and Llama 3.2 3B, which are optimized for edge and mobile devices; and Llama Guard 3 11B Vision, which is fine-tuned for content safety classification, according to the update. "According to Meta, Llama 3.2 models have been evaluated on over 150 benchmark datasets, demonstrating competitive performance with leading foundation models," the update said. "Similar to Llama 3.1, all of the models support a 128K context length and multilingual dialogue use cases across eight languages, spanning English, German, French, Italian, Portuguese, Hindi, Spanish and Thai." Meta announced Llama 3.2, a big advancement in its open-source AI model series, on Wednesday at its annual Connect conference. The vision models can analyze images, understand charts and graphics, and perform visual grounding tasks, PYMNTS reported Wednesday. The lightweight models, optimized for on-device use, support multilingual text generation and tool-calling abilities, enabling developers to build personalized applications supposedly prioritizing user privacy. "It's only been a year and a half since we first announced Llama, and we've made incredible progress in such a short amount of time," Meta said in a Wednesday blog post. "This year, Llama has achieved 10x growth and become the standard for responsible innovation. Llama also continues to lead on openness, modifiability and cost efficiency, and it's competitive with closed models -- even leading in some areas."

Meta's Llama AI models now support images, too | TechCrunch

Benjamin Franklin once wrote that nothing is certain except death and taxes. Let me amend that phrase to reflect the current AI gold rush: Nothing is certain except death, taxes, and new AI models, with the last of those three arriving at an ever-accelerating pace. Earlier this week, Google released upgraded Gemini models, and, earlier in the month, OpenAI unveiled its o1 model. But on Wednesday, it was Meta's turn to trot out its latest at the company's annual Meta Connect 2024 developer conference in Menlo Park. Meta's multilingual Llama family of models has reached version 3.2, with the bump from 3.1 signifying that several Llama models are now multimodal. Llama 3.2 11B -- a compact model -- and 90B, which is a larger, more capable model, can interpret charts and graphs, caption images, and pinpoint objects in pictures given a simple description. Given a map of a park, for example, Llama 3.2 11B and 90B might be able to answer questions like, "When will the terrain become steeper?" and "What's the distance of this path?" Or, provided a graph showing a company's revenue over the course of a year, the models could quickly spotlight the best-performing months of the bunch. For developers who wish to use the models strictly for text applications, Meta says that Llama 3.2 11B and 90B were designed to be "drop-in" replacements for 3.1. 11B and 90B can be deployed with or without a new safety tool, Llama Guard Vision, that's designed to detect potentially harmful (i.e. biased or toxic) text and images fed to or generated by the models. In most of the world, the multimodal Llama models can be downloaded from and used across a wide number of cloud platforms, including Hugging Face, Microsoft Azure, Google Cloud, and AWS. Meta's also hosting them on the official Llama site, Llama.com, and using them to power its AI assistant, Meta AI, across WhatsApp, Instagram, and Facebook. But Llama 3.2 11B and 90B can't be accessed in Europe. As a result, several Meta AI features available elsewhere, like image analysis, are disabled for European users. Meta once again blamed the "unpredictable" nature of the bloc's regulatory environment. Meta has expressed concerns about -- and spurned a voluntary safety pledge related to -- the AI Act, the EU law that establishes a legal and regulatory framework for AI. Among other requirements, the AI Act mandates that companies developing AI in the EU commit to charting whether their models are likely to be deployed in "high-risk" situations, like policing. Meta fears that the "open" nature of its models, which give it little insight into how the models are being used, could make it challenging to adhere to the AI Act's rules. Also at issue for Meta are provisions in the GDPR, the EU's broad privacy law, pertaining to AI training. Meta trains models on the public data of Instagram and Facebook users who haven't opted out -- data that in Europe is subject to GDPR guarantees. EU regulators earlier this year requested that Meta halt training on European user data while they assessed the company's GDPR compliance. Meta relented, while at the same time endorsing an open letter calling for "a modern interpretation" of GDPR that doesn't "reject progress." Earlier this month, Meta said that it would resume training on U.K. user data after "[incorporating] regulatory feedback" into a revised opt-out process. But the company has yet to share an update on its training throughout the rest of the bloc. Other new Llama models -- models that weren't trained on European user data -- are launching in Europe (and globally) Wednesday. Llama 3.2 1B and 3B, two lightweight, text-only models designed to run on smartphones and other edge devices, can be applied to tasks such as summarizing and rewriting paragraphs (e.g. in an email). Optimized for Arm hardware from Qualcomm and MediaTek, 1B and 3B can also tap tools such as calendar apps with a bit of configuration, Meta says, allowing them to take actions autonomously. There isn't a follow-up, multimodal or not, to the flagship Llama 3.1 405B model released in August. Given 405B's massive size -- it took months to train -- it's likely a matter of constrained compute resources. We've asked Meta if there are other factors at play and will update this story if we hear back. Meta's new Llama Stack, a suit of Llama-focused dev tools, can be used to fine-tune all the Llama 3.2 models: 1B, 3B, 11B, and 90B. Regardless of how they're customized, the models can process up to around 100,000 words at once, Meta says. Meta CEO Mark Zuckerberg often talks about ensuring all people have access to the "benefits and opportunities" of AI. Implicit in this rhetoric, however, is a desire that these tools and models be of Meta's making. Spending on models that it can then commoditize forces the competition (e.g. OpenAI, Anthropic) to lower prices, spreads Meta's version of AI broadly, and lets Meta incorporate improvements from the open source community. Meta claims that its Llama models have been downloaded over 350 million times and are in use by large enterprises including Zoom, AT&T, and Goldman Sachs. For many of these developers and companies, it's immaterial that the Llama models aren't "open" in the strictest sense. Meta's license constrains how certain devs can use them; platforms with over 700 million monthly users must request a special license from Meta that the company will grant on its discretion. Granted, there aren't many platforms of that size without their own in-house models. But Meta isn't being especially transparent about the process. When I asked the company this month whether it had approved a discretionary Llama license for a platform yet, a spokesperson told me that Meta "didn't have anything to share on the topic." Make no mistake, Meta's playing for keeps. It's spending millions lobbying regulators to come around to its preferred flavor of "open" AI, and it's ploughing billions into servers, datacenters, and network infrastructure to train future models. None of the Llama 3.2 models solve the overriding problems with today's AI, like its tendency to make things up and regurgitate problematic training data (e.g. copyrighted e-books that might've been used without permission, the subject of a class action lawsuit against Meta). But, as I've written before, they do advance one of Meta's key goals: becoming synonymous with AI, and in particular generative AI.

AI for all: Meta's 'Llama Stack' promises to simplify enterprise adoption

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Today at its annual Meta Connect developer conference, Meta launched Llama Stack distributions, a comprehensive suite of tools designed to simplify AI deployment across a wide range of computing environments. This move, announced alongside the release of the new Llama 3.2 models, represents a major step in making advanced AI capabilities more accessible and practical for businesses of all sizes. The Llama Stack introduces a standardized API for model customization and deployment, addressing one of the most pressing challenges in enterprise AI adoption: the complexity of integrating AI systems into existing IT infrastructures. By providing a unified interface for tasks such as fine-tuning, synthetic data generation, and agentic application building, Meta positions Llama Stack as a turnkey solution for organizations looking to leverage AI without extensive in-house expertise. Cloud partnerships expand Llama's reach Central to this initiative is Meta's collaboration with major cloud providers and technology firms. Partnerships with AWS, Databricks, Dell Technologies, and others ensure that Llama Stack distributions will be available across a wide range of platforms, from on-premises data centers to public clouds. This multi-platform approach could prove particularly attractive to enterprises with hybrid or multi-cloud strategies, offering flexibility in how and where AI workloads are run. The introduction of Llama Stack comes at a critical juncture in the AI industry. As businesses increasingly recognize the potential of generative AI to transform operations, many have struggled with the technical complexities and resource requirements of deploying large language models. Meta's approach, which includes both powerful cloud-based models and lightweight versions suitable for edge devices, addresses the full spectrum of enterprise AI needs. Breaking down barriers to AI adoption The implications for IT decision-makers are substantial. Organizations that have been hesitant to invest in AI due to concerns about vendor lock-in or the need for specialized infrastructure may find Llama Stack's open and flexible approach compelling. The ability to run models on-device or in the cloud using the same API could enable more sophisticated AI strategies that balance performance, cost, and data privacy considerations. However, Meta's initiative faces challenges. The company must convince enterprises of the long-term viability of its open-source approach in a market dominated by proprietary solutions. Additionally, concerns about data privacy and model safety need addressing, particularly for industries handling sensitive information. Meta has emphasized its commitment to responsible AI development, including the release of Llama Guard 3, a safeguard system designed to filter potentially harmful content in both text and image inputs. This focus on safety could be crucial in winning over cautious enterprise adopters. The future of enterprise AI: Flexibility and accessibility As enterprises evaluate their AI strategies, Llama Stack's promise of simplified deployment and cross-platform compatibility is likely to attract significant attention. While it's too early to declare it the de facto standard for enterprise AI development, Meta's bold move has undoubtedly disrupted the competitive landscape of AI infrastructure solutions. The real strength of Llama Stack is its ability to make AI development more accessible to businesses of all sizes. By simplifying the technical challenges and reducing the resources needed for AI implementation, Meta is opening the door for widespread innovation across industries. Smaller companies and startups, previously priced out of advanced AI capabilities, might now have the tools to compete with larger, resource-rich corporations. Moreover, the flexibility offered by Llama Stack could lead to more nuanced and efficient AI strategies. Companies might deploy lightweight models on edge devices for real-time processing, while leveraging more powerful cloud-based models for complex analytics -- all using the same underlying framework. For business and tech leaders, Llama Stack offers a simpler path to using AI across their companies. The question is no longer if they should use AI, but how to best fit it into their current systems. Meta's new tools could speed up this process for many industries. As companies rush to adopt these new AI capabilities, one thing is clear: the race to harness AI's potential is no longer just for tech giants. With Llama Stack, even the corner store might soon be powered by AI.

Meta Releases Llama 3.2 -- and Gives Its AI a Voice

Meta's AI assistants can now talk and see the world. The company is also releasing the multimodal Llama 3.2, a free model with visual skills. Mark Zuckerberg announced today that Meta, his social-media-turned-metaverse-turned-artificial intelligence conglomerate, will upgrade its AI assistants to give them a range of celebrity voices, including those of Dame Judi Dench and John Cena. The more important upgrade for Meta's long-term ambitions, though, is the new ability of its models to see users' photos and other visual information. Meta today also announced Llama 3.2, the first version of its free AI models to have visual abilities, broadening their usefulness and relevance for robotics, virtual reality, and so-called AI agents. Some versions of Llama 3.2 are also the first to be optimized to run on mobile devices. This could help developers create AI-powered apps that run on a smartphone and tap into its camera or watch the screen in order to use apps on your behalf. Given Meta's enormous reach with Facebook, Instagram, WhatsApp, and Messenger, the assistant upgrade could give many people their first taste of a new generation of more vocal and visually capable AI helpers. Meta said today that more than 180 million people already use Meta AI, as the company's AI assistant is called, every week. Meta has lately given its AI a more prominent billing in its apps -- for example, making it part of the search bar in Instagram and Messenger. The new celebrity voice options available to users will also include Awkwafina, Keegan Michael Key, and Kristen Bell. Meta previously gave celebrity personas to text-based assistants, but these characters failed to gain much traction. In July the company launched a tool called AI Studio that lets users create chatbots with any persona they choose. Meta says the new voices will be made available to users in the US, Canada, Australia, and New Zealand over the next month. The Meta AI image capabilities will be rolled out in the US, but the company did not say when the features might appear in other markets. The new version of Meta AI will also be able to provide feedback on and information about users' photos; for example, if you're unsure what bird you've snapped a picture of, it can tell you the species. And it will be able to help edit images by, for instance, adding new backgrounds or details on demand. Google released a similar tool for its Pixel smartphones and for Google Photos in April. Powering Meta AI's new capabilities is an upgraded version of Llama, Meta's premier large language model. The free model announced today may also have a broad impact, given how widely the Llama family has been adopted by developers and startups already. In contrast to OpenAI's models, Llama can be downloaded and run locally without charge -- although there are some restrictions on large-scale commercial use. Llama can also more easily be fine-tuned, or modified with additional training, for specific tasks. Patrick Wendell, cofounder and VP of engineering at Databricks, a company that hosts AI models including Llama, says many companies are drawn to open models because they allow them to better protect their own data.