Thinking Machines Lab releases Inkling, a massive open-weight AI model to challenge closed systems

Thinking Machines amps up its bet against one-size-fits-all AI with its first open model, Inkling

Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, released its first proprietary AI model Wednesday morning, called Inkling -- and unlike the flagship models from OpenAI, Anthropic, or Google, it's open-weight, meaning outside developers and companies can download it and modify it directly. Inkling is a mixture-of-experts system with 975 billion total parameters, though it only draws on a fraction of that -- about 41 billion -- for any given task, a common design that keeps very large models faster and cheaper to run. It was trained on 45 trillion tokens of text, image, audio, and video, and reasons natively across all three, according to the company's own release materials. It's the company's first public proof point after a year and a half spent building AI infrastructure largely out of public view. Some of that work surfaced already, in a May research preview of "interaction models" -- AI designed to listen and speak (and even interrupt) instead of stop and wait as with typical chatbots. It's also a test of the central bet behind Thinking Machines, which is that AI that organizations can adapt for themselves will outperform the one-size-fits-all models the biggest labs currently sell. It's an interesting model, one that's designed to give calibrated answers, including flagging uncertainty rather than guessing, and which lets users dial "thinking effort" up or down when they want to trade for speed. On one benchmark, the company says, Inkling uses a third as many tokens as Nvidia's Nemotron 3 Ultra in order to hit the same coding performance. It's worth noting that Thinking Machines doesn't claim Inkling is best-in-class. Its briefing materials state explicitly that Inkling is "not the strongest model available today, closed or open." What it's evidently going for instead is well-rounded performance. Of course, that raises a big question, which is who this product is targeting, beyond the obvious -- this is definitely an enterprise product. Thinking Machines is, for now, marketing it less as a finished work than as a starting point, something for organizations to fine-tune themselves through Tinker, the company's model-customization platform. (OpenAI, Anthropic, and Google have all taken a very different approach with ChatGPT, Claude, and Gemini, respectively, which were all built to compete as general-purpose chatbots first, with agentic, autonomous features layered on top.) A post published by Thinking Machines last week was clearly meant as the backdrop for this release. AI that's trained centrally by one company and then set in stone, the company argued in that post, underperforms AI that organizations shape themselves because so much expertise is specific to the people who hold it. The broader idea is that centralized labs are selling everyone the same product, repeatedly refined by the lab that built it, while enterprises willing to own and customize their own models can wring far more value from them. It's an argument that's gaining steam. In a blog post published Sunday, Microsoft CEO Satya Nadella -- whose company has invested billions in both OpenAI and Anthropic -- warned that enterprises using proprietary AI models effectively pay twice: once in subscription costs, and again by handing over business knowledge embedded in their thousands of prompts and corrections, which can be absorbed into future model versions. Hugging Face CEO Clem Delangue made a similar prediction in conversation with TechCrunch last week. Frontier models, he said, will increasingly be reserved for experimentation and high-value tasks, while most production AI work shifts to private or open-source alternatives -- the exact split Thinking Machines is building around. The clearest evidence for that argument came recently from a project with Bridgewater Associates, the world's largest hedge fund (which is not, for what it's worth, a Thinking Machines investor). Researchers from both companies took an existing open-source model and trained it further on Bridgewater's own financial expertise. The result scored 84.7% on financial reasoning tests, beating top proprietary AI models, while costing roughly a fourteenth as much to run, though those results, published jointly in late June, come from the two companies' own evaluation, not an independent one. Thinking Machines has also emphasized how quickly it got here: OpenAI took roughly five years, and Anthropic roughly three, to bring tech to market and show revenue; Thinking Machines says it did the same in about nine months. Some will wonder whether Inkling was trained on outputs from competitors' models, a practice known as distillation that has drawn scrutiny industry-wide. The short answer, per the company's own materials, is partly. Thinking Machines pretrained Inkling from scratch, but it says it used other open-weight models -- including Moonshot AI's Kimi K2.5 -- to help generate some of its early post-training data before large-scale reinforcement learning took over. The next model, the company insists, will use fully self-contained post-training instead. On the cost side, Thinking Machines has been more guarded. It struck a strategic partnership with Nvidia in March to deploy a gigawatt of Vera Rubin computing capacity, and says Inkling itself was trained entirely on Nvidia's GB300 NVL72 systems. But the company hasn't said how it plans to balance that against revenue that, by most accounts, hasn't been a primary focus so far. (A reported $50 billion fundraising round was said to be coming together last November, which multiple outlets reported had stalled by January; the company has declined to talk about its funding picture since, though Nvidia said it made a "significant investment" in Thinking Machines when the companies announced that March partnership.) A related question is whether Thinking Machines' spending will ever reach the scale of OpenAI's or Anthropic's, or whether its efficiency-driven approach means the economics look different. Put another way, the company's bet may be less that it will eventually spend like its larger rivals than that it won't need to at all -- because once weights are public, nothing obligates anyone who downloads them to pay Thinking Machines to run them, unlike the metered access OpenAI and Anthropic sell. It's Tinker, not the model itself, where the company's revenue has to come from, via training, fine-tuning, and, now, a cut of the hosting ecosystem built around it.

Thinking Machines Lab Drops Its First Model

Thinking Machines Lab, an artificial intelligence company started by exiles from OpenAI, has released its first model, called Inkling. The startup's new model is open-weight, which means that researchers and startups will be able to download and modify it. In a a blog post, the company says Inkling was trained from scratch to make sense of audio and video input as well as text. It says that while Inkling isn't the best model on popular benchmarks, it performs well at many tasks, and is capable of advanced reasoning and coding. Like many open-weight models, Inkling is relatively large -- 975 billion parameters -- and needs to run on a cluster of specialized chips. In a sign of how AI models are increasingly being used to build AI, the lab also used Inkling to fine-tune and improve itself. The release could help Thinking Machines establish itself as a legitimate player in the frenetic and big-spending AI race. Open-source models have proven popular because they're cheaper to run than closed models, which can typically only be accessed for a fee. Open-source models can also be more easily modified for different tasks. The best open-weight models currently come from China, but Thinking Machines says Inkling offers a level of performance similar to those models. The release of an open-weight model fits with a vision for AI that Thinking Machines laid out in a recent blog post. The company said the technology shouldn't be controlled by just a few companies and should be decentralized so that more people can build their own models with their own data. According to a company source who requested anonymity to discuss the development process, researchers discovered a strange phenomenon while training Inkling. Like other models, it usually provides a natural language explanation for its complex reasoning. Inkling decided to do away with this in the name of efficiency. "It determined that the grammar was overhead, which is interesting," the source says. The company reinstated natural language reasoning to make the models' decisions more explainable, the person says. Thinking Machines was founded in February 2025 by several big-name executives and researchers from OpenAI, including Mira Murati, who served as CTO (and briefly CEO) of OpenAI; John Schulman, a cofounder of OpenAI who played a key role in developing ChatGPT; and Lilian Weng, a former VP at OpenAI who led work on safety and robotics. The startup received the largest seed funding round in history, which valued it at $12 billion out of the gate. Previously, the company released Tinker, a tool for fine-tuning models, showcased a tool that enables natural voice interactions, and published machine-learning research. OpenAI may have kick-started the AI boom with ChatGPT, but defector-led companies like Thinking Machines and Anthropic have muscled into the space. Anthropic recently filed for an IPO, and the company is valued at more than a trillion dollars. Its model Claude has proven popular with many businesses, especially for its coding skills.

Mira Murati's Thinking Machines Lab has debuted its first AI model.

Murati, OpenAI's former CTO (and, briefly, CEO during Sam Altman's ouster in 2023), wrote on X that the open-weight model, called "Inkling," was trained from scratch. Judging by the official blog post, though, it seemed like the company was purposefully setting expectations for Inkling low: "It is not the most performant model available today, closed or open ... We trained Inkling for solid capabilities across the board rather than state-of-the-art performance in a single area, to serve as a foundation for the models we will train in the future."

Former OpenAI CTO does what Altman won't: releases a frontier AI model that's actually open

If you're in the market for a frontier-class open weights model, your options are few and far between outside of the Chinese model houses. With the Wednesday release of a new model code-named "Inkling", an outfit called Thinking Machines Lab aims to change that. Founded in early 2025 by former OpenAI CTO Mira Murati, Thinking Machines' first model is a big one. Weighing in at 975 billion parameters, the model requires more than two terabytes of GPU memory -- a quantity present in around eight of Nvidia's B300 accelerators, or sixteen H200s -- to run at its native 16-bit precision. If that's asking too much of your hardware, Thinking Machines has also released a NVFP4 quantized version of the model capable of running on half the GPUs. This makes it the largest American open weights model to date, and comparable to Chinese models like DeepSeek V4, GLM 5.2, and Kimi K2.6 in terms of size and capabilities. Take these claims with a grain of salt -- gaming AI benchmarks isn't exactly difficult -- but Thinking Machines says Inkling is competitive with these models in a variety of workloads, although its benchmark charts also show it trailing proprietary models like Anthropic's Claude and OpenAI's GPT. Thinking Machines describes the model as being highly adaptable, intended for use by developers building AI apps, but suitable for general purpose applications like chat bots. And because it's being released under a highly permissive Apache 2.0 license, end users are free to fine tune it for their specific use case. The company's Tinker platform offers tools to do just that. In fact, Thinking Machines boasts that the model is capable of writing its own fine tuning scripts to refine its behavior, teach itself new skills, and evaluate its abilities. Other notable features include support for a million-token context, which you can think of as the model's short-term memory. This should help it wrangle large code bases and needle-in-the-haystack type search problems. While Thinking Machines admits the model's mixture of experts (MoE) architecture was inspired by DeepSeek-V3, the company says it trained Inkling from scratch using Nvidia GB300 NVL72 systems and 45 trillion tokens worth of text, images, audio, and video. In total, the model features 256 routed exports and two shared ones. The model generates each token by six experts, totaling about 41 billion parameters. So, in spite of its size, the model should be able to churn out tokens at about the same rate as DeepSeek V4 when running on the same hardware. Like most LLMs today, Inkling is a "reasoning model" which is to say it's been trained using reinforcement learning (RL) to use chain of thought to "think" through requests before responding. The model developer claims to have tuned the model to use these thinking tokens more efficiently and that Inkling therefore matches Nvidia's Nemotron 3 Ultra, up to now the largest and most capable American open weights model out there at 550 billion parameters, on Terminal Bench 2.1 using roughly a third the tokens. Thinking tokens may make models more capable and less likely to hallucinate, but the capability comes at a cost. Those tokens are billed like any other and so the longer the model thinks, the larger users' bills become. Speaking of APIs, Inkling is available starting today on Thinking Machines' Tinker platform, which in addition to model access also offers tools for customization and fine tuning. The company is also working to bring the model to 3rd-party API services including TogetherAI, Fireworks, Modal, Databricks, and Baseten. If you prefer to evaluate the model on your own hardware, it's available for download on popular model repos like Hugging Face. At launch, the model claims support for a broad range of inference engines including vLLM, SGLang, Miles, TokenSpeed, and Llama.cpp. Inkling is the first of several new models under development by Thinking Machines. Alongside its flagship model, the company is also previewing Inkling-Small, a 276-billion-parameter MoE model with 12 billion active parameters for those prioritizing latency over throughput and quality. Thinking Machines -- which shares its name with the fictional supercomputer maker immortalized in 1993's Jurassic Park -- is currently in the process of finalizing the model and plans to release its weights once testing is complete. ®

AI startup Thinking Machines launches an open-weight AI model

SAN FRANCISCO, July 15 (Reuters) - AI startup Thinking Machines revealed on Wednesday a new artificial intelligence model that could serve as one of the few alternatives to popular open-source offerings from Chinese AI labs. Named Inkling, the model is open-weight, meaning users can download, run and customize the underlying systems, unlike proprietary, closed-source models. It is the first general-purpose model release to come out of Thinking Machines, a San Francisco-based startup founded last year by OpenAI's former chief technology ⁠officer Mira Murati. Thinking Machines launched its first product called Tinker, which helps customize AI models, last October. Inkling is available on Tinker and other developer platforms, it said. The model has 975 billion parameters -- variables that determine how an AI system processes information -- making it one of the largest models of its kind. The open-source ecosystem in the West lags behind its counterpart in China, especially in the wake of a void left by Meta (META.O), opens new tab, which changed ⁠course to a proprietary approach after the disappointing release of its open Llama 4 model last year. Businesses have in turn flocked to adopt Chinese models as the primary alternatives to expensive closed-source models. Hedge fund Bridgewater Associates used Tinker to build a ⁠custom version of Qwen, a model developed by China's Alibaba (9988.HK), opens new tab, which it said outperformed top proprietary models at lower costs. Thinking Machines published a series of benchmarks that compared ⁠Inkling's capabilities with closed models from Anthropic, Google (GOOGL.O), opens new tab, and OpenAI, as well as leading open offerings, most of them from Chinese labs. While those ⁠other models maintain the edge on performance overall, Inkling put in a competitive showing, particularly on agent-related tasks, that could spur interest from prospective users. Reporting by Kenrick Cai in San Francisco; Editing by Nia Williams Our Standards: The Thomson Reuters Trust Principles., opens new tab * Suggested Topics: * Artificial Intelligence Kenrick Cai Thomson Reuters Kenrick Cai is a correspondent for Reuters based in San Francisco. He covers Google, its parent company Alphabet and artificial intelligence. Cai joined Reuters in 2024. He previously worked at Forbes magazine, where he was a staff writer covering venture capital and startups. He received a Best in Business award from the Society for Advancing Business Editing and Writing in 2023. He is a graduate of Duke University. Reach him on Signal at @kenrick.01.

Mira Murati's Thinking Machines draws from Chinese rivals in debut AI model

Thinking Machines Lab has unveiled its first general-purpose AI model, with OpenAI's former chief technology officer Mira Murati leaning on cheap Chinese technology to develop its products. The San Francisco-based "neo lab" on Wednesday said its Inkling foundation model's architecture drew on China's DeepSeek-V3 and has been refined post-training using data generated by Beijing-based Moonshot AI's Kimi K2.5. Its release comes at a time when China's AI labs are gaining ground on US rivals. Chinese AI models have overtaken US counterparts in overall usage this year, according to data platform OpenRouter. Leading US labs have repeatedly complained about Chinese rivals copying their models, including through so-called distillation -- in which outputs from a larger "teacher" model train a smaller "student" model. Thinking Machines' Inkling ranks below the leading products from Anthropic and OpenAI as well as several Chinese models, according to benchmark data shared by the company. But in contrast to the top ChatGPT and Claude models, Inkling will be open-weight, meaning users will be able to host the model on their servers and fine-tune it for specific use cases. "Today we are advancing our mission by releasing a model we trained from scratch with the full weights available, so that people can make it their own," the company said. The debut model comes a year after the Thinking Machines closed a $2bn seed fundraising round at a $12bn post-money valuation, having received backing from Andreessen Horowitz alongside chipmakers Nvidia and AMD as well as hedge fund Jane Street. The start-up has attracted attention thanks in part to Murati's status in the AI industry, having worked on the development of several OpenAI products including ChatGPT, the Dall-E image generator and its voice mode. She also played a leading role in the brief ousting of OpenAI chief executive Sam Altman during a board coup in November 2023. The former OpenAI CTO departed in September 2024 before founding Thinking Machines in February last year. Thinking Machines suffered a setback earlier this year when several senior staff members departed for Meta and OpenAI and questioned Murati's leadership. The start-up has moved forward with several products including its Tinker platform, which allows enterprise customers to fine-tune and customise large language models for their particular business applications. Thinking Machines in a blog post on Friday said that most AI tools were "trained in a handful of places and then frozen" rather than being shaped by users. "Extending human will and judgment calls for AIs as diverse and distributed as people themselves are," the company said. Rivals including OpenAI offer open-weight models that are similarly customisable, while platforms such as Amazon and Microsoft boast services that enable users to fine-tune models using their own datasets. Murati's AI lab has also sought to differentiate itself by developing services that work in tandem with people in their daily lives. In May, the start-up previewed its first "interaction model", which can take actions based on users' audio and video inputs rather than exclusively interpreting text.

Thinking Machines Lab offers enterprises a US alternative in open-weight AI

Inkling, a 975-billion-parameter model, can be customized through Tinker and supports a 1-million-token context window, but it enters a market where Chinese models lead several coding and reasoning benchmarks. Thinking Machines Lab, the San Francisco startup founded by former OpenAI CTO Mira Murati, has released Inkling, its first general-purpose AI model. The launch adds another US-developed entrant to an open-weight market where Chinese developers produce several leading coding and reasoning models. Inkling uses a mixture-of-experts architecture with 975 billion total parameters, of which 41 billion are active during processing. It supports a context window of up to 1 million tokens and was pretrained on 45 trillion tokens spanning text, images, audio, and video. Thinking Machines said it also trained the model for coding, tool use, and multimodal tasks. The release follows the October 2025 launch of Tinker, Thinking Machines' first product and an API-based platform for customizing AI models. Developers can fine-tune Inkling through the platform.

Thinking Machines Inkling: Murati's first open-weight model

Mira Murati's lab has released Thinking Machines Inkling, its first model. It is open-weight, enormous, and, by the company's own admission, not the best one out there. "We believe in keeping the weirdness alive." That line comes from a manifesto Mira Murati's lab published last week. It is also the thinking behind the lab's first model. Thinking Machines Lab, founded by the former OpenAI chief technology officer, has released Inkling. It is open-weight, so any developer or company can download the model and reshape it. That alone sets it apart from the flagships sold by OpenAI, Anthropic, and Google. Inkling is big. It is a mixture-of-experts system with 975 billion total parameters, though it uses only about 41 billion for any given task. It handles a context window of up to 1 million tokens, and it trained on 45 trillion tokens of text, images, audio, and video. It reasons across text, images, and audio, but for now it only writes text back, including code and structured data. A model that admits it is not the best Here is the twist. Thinking Machines does not claim Inkling tops the charts. Its own materials call it "not the strongest model available today, closed or open." The lab is chasing something else: range and adaptability. Thinking Machines Inkling is meant to be a broad, balanced base that organisations fine-tune for their own work, not a finished chatbot. Users can dial its "thinking effort" up or down to trade accuracy for speed. On one coding test, the company says, Inkling matches Nvidia's Nemotron 3 Ultra using a third as many tokens. The lab also previewed a lighter model, Inkling-Small, with 12 billion active parameters. It suits jobs where cost and speed matter most. The bet: shape it yourself The whole release rests on one wager. AI trained in one place and then frozen, the lab argues, loses to AI that each organisation can shape around its own expertise. Customers fine-tune Inkling through Tinker, Thinking Machines' customisation platform, and they own the result. They also carry the safety risk of whatever they build. The lab points to a project with the hedge fund Bridgewater as proof. The two trained an open model on Bridgewater's financial know-how, and it scored 84.7% on financial reasoning tests, beating top proprietary models at a fraction of the cost. That figure comes from the two companies' own evaluation, not an independent one. The argument is gaining company. Microsoft's Satya Nadella recently warned that firms using closed models pay twice, once in fees and once by handing over the knowledge baked into their prompts. Cheap open-weight models, many from China, pull the same way. Nine months, with some borrowed help Thinking Machines is keen to stress its speed. OpenAI took about five years to ship and earn, and Anthropic roughly three, TechCrunch noted. Murati's lab says it did it in about nine months. It cut a few corners to get there. To start Inkling's training, the lab leaned on other open models, including Moonshot's Kimi K2.5, a practice known as distillation. Its next model, it insists, will train fully on its own. Inkling ran on Nvidia's GB300 systems, part of a March deal for a gigawatt of Nvidia compute. Money and people have been bumpier. The lab raised $2bn at a $12bn valuation last year, and a reported $50bn round stalled. Two co-founders left earlier this year, though headcount is back to around 200. For now, Thinking Machines will not charge for Inkling at all. Its money comes from Tinker, and its case rests on the weirdness holding up.

Thinking Machines' first model bets big on customization

Why it matters: Thinking Machines is making a different bet than many AI labs: that enterprises ultimately care less about the smartest general-purpose model than one they can make their own. State of play: Inkling is the first foundational model release from Thinking Machines. * The company is clear that Inkling is not the strongest model available, instead focusing on how the model is customizable, which could help users get better performance with lower costs. * Instead of starting with another company's AI and modifying from there, this model was built from scratch and trained from the ground up. * Thinking Machines trained the model on Nvidia's latest AI infrastructure, underscoring the company's partnership with the chip giant. The intrigue: Thinking Machines used data generated by existing open models -- including Kimi K2.5 from Chinese lab Moonshot AI -- in its final training phase. Between the lines: Instead of trying to beat competitors like OpenAI and Anthropic on model benchmarks, Thinking Machines is currently focused on customization. * Still, Inkling could be the company's first step towards more powerful successors that Thinking Machines is already training. * Thinking Machines is also previewing Inkling-Small, a lighter-weight model with weights that will be released after testing. Zoom in: The full weights are available on Hugging Face and the model is now live for fine-tuning on Tinker, the company's customization platform. Zoom out: Demand for open-weight models is growing as companies look for cheaper AI they can customize for their own applications. * Organizations can fine tune open-weight models on their own proprietary data and deploy those models on infrastructure that they control, giving users more flexibility over hosting and costs than most closed models. * Palantir CEO Alex Karp recently went viral discussing this dynamic on CNBC, saying frontier AI tools from closed model providers are too expensive and don't offer enough clarity on IP protections. Flashback: This isn't Murati's first go-round with the open/closed model dilemma. * She was at OpenAI in 2019 when the lab -- founded on a promise of openness -- withheld the full version of GPT-2 over misuse fears, heralding the company's retreat from fully open releases. * Just because Inkling is open weight, that doesn't mean the rest of Thinking Machines' models will be. Her current strategy could mirror that case-by-case logic: release models openly when the risks are manageable and hold them back when they aren't. Follow the money: Thinking Machines raised a record $2 billion seed round at a $12 billion valuation in 2025, before it had released a model or product. * Nvidia was among the startup's investors and has since deepened the relationship. * Thinking Machines also reportedly signed a multibillion-dollar Google Cloud deal. The bottom line: Inkling is Thinking Machines' first real test of whether its billions in funding can translate into a compelling alternative for enterprises looking to customize AI.

Thinking Machines open sources first multimodal language model, Inkling, focused on low cost and 'resistance to censorship'

Enterprises looking to move more of their agentic AI workloads to open weights models they can customize, control and run on-premises or in virtual private clouds have a strong new contender to consider. Today, Thinking Machines -- the highly capitalized American AI startup founded by former OpenAI CTO Mira Murati -- released Inkling, its first major language model under an enterprise-friendly Apache 2.0 open source license, and it boasts high, if sub state-of-the-art, performance for open weights models on third-party benchmarks, specifically software engineering (77.6% on SWE-bench Verified, where it beats fellow U.S. open rival Nvidia Nemotron 3's 71.9%) and voice understanding (91.4% on VoiceBench compared to 94.4% for Gemini 3.1 Pro on high reasoning effort). Another differentiator: Thinking Machines notes that Inkling was designed "to answer directly on topics that may be subject to censorship," offering enterprises concerned about factual outputs, irrespective of controversy or sensitivity, a more trustworthy option. Coming in at 975 billion total parameters, Inkling is a natively multimodal, open-weights Mixture-of-Experts (MoE) system capable of reasoning across text, images, and audio. The weights are already available on Hugging Face and the company's own model training application programming interface (API), Tinker. Designed to balance cost against performance through a novel "controllable thinking effort" mechanism, the model represents a significant departure from the black-box scaling strategies of frontier competitors. Alongside the flagship model, Thinking Machines also announced a preview of Inkling-Small, a lighter 276-billion-parameter alternative optimized for workloads where low latency and cost are paramount. Benchmarks Show a Powerful, High-End, Sub State-of-the-Art Model While Inkling is a formidable multimodal engine, it lands in a fiercely competitive 2026 open-weight landscape characterized by highly specialized MoE architectures. Rather than attempting to dominate every leaderboard, Thinking Machines explicitly designed Inkling -- with 975 billion total and 41 billion active parameters -- as a broad, balanced generalist. For example, it comes in near the middle high-end of benchmark performance 1257 on Design Arena's Agentic Web Dev leaderboard measuring human scores of frontend web design. But China's leading AI labs have produced models with elite reasoning and coding capabilities, posing a stiff challenge to Inkling's generalist approach and ultimately outperforming it on general and coding benchmarks. * GLM 5.2: Widely considered the top open-weight reasoning model available in the benchmark set, GLM 5.2 outperforms Inkling on pure coding, agentic, and complex reasoning tasks. It scores 62.1% on SWEBench Pro (Public) compared to Inkling's 54.3%, and a massive 82.7 on Terminal Bench 2.1 against Inkling's 63.8. GLM 5.2 also holds the edge in text-only reasoning, scoring 40.1% on HLE (text only) versus Inkling's 30.0%. * DeepSeek V4 Pro: DeepSeek maintains an edge in several strict coding and factuality domains, beating Inkling on SWEBench Verified (80.6% vs. 77.6%) and SimpleQA Verified (57.0% vs. 43.9%). However, Inkling successfully overtakes DeepSeek V4 Pro in mathematical problem-solving, achieving 97.1% on AIME 2026 compared to DeepSeek's 96.7%. * Kimi K2.6: This model outpaces Inkling across multiple technical benchmarks, delivering higher scores on GPQA Diamond (91.1% vs. 87.9%), BrowseComp (83.2% vs. 77.1%), and HLE with tools (54.0% vs. 46.0%). Yet Inkling proves more resilient on general chat instruction following, scoring 79.8% on IFBench compared to Kimi K2.6's 76.0%. Against its primary U.S.-based open-weight competition, Inkling demonstrates strong parity and frequent superiority. * Nemotron 3 Ultra: Inkling consistently outperforms this U.S. rival across reasoning and coding. Inkling posts 97.1% on AIME 2026 and 77.6% on SWEBench Verified, beating Nemotron's 94.2% and 70.7%, respectively. Furthermore, Inkling significantly leads in agentic workflows, scoring 74.1% on MCP Atlas against Nemotron's 44.7%. When compared to closed-source juggernauts like Claude Fable 5, GPT 5.6 Sol, and Gemini 3.1 Pro, Inkling trails in peak reasoning and software engineering autonomy, but remains highly competitive in multimodality. * Coding and Reasoning: Closed models maintain a commanding lead. Claude Fable 5 (max) hits 95.0% on SWEBench Verified and 53.3% on HLE (text only), far outpacing Inkling's 77.6% and 30.0%. GPT 5.6 Sol dominates Terminal Bench 2.1 with an 89.5, easily clearing Inkling's 63.8. * Native Multimodality: Inkling's native visual and audio capabilities hold their own. On the MMMU Pro (Standard 10) vision benchmark, Inkling's 73.3% is competitive, though trailing Claude Fable 5's 84.2% and GPT 5.6 Sol's 83.0%. In audio processing, Inkling scores a highly respectable 77.2% on MMAU, keeping it within striking distance of Gemini 3.1 Pro's 82.5%. If an enterprise workflow demands elite software engineering autonomy or the highest bounds of text-only reasoning, models like GLM 5.2 or proprietary systems like Claude Fable 5 maintain the edge. However, Inkling carves out a unique and highly defensible position: it is the most capable open-weight foundation model that natively fuses text, vision, and audio, while simultaneously offering developers direct programmatic control over the cost-to-performance ratio. The Shift from Static Reasoning to Controllable Thinking Rather than attempting to build a singular "god model" optimized strictly for state-of-the-art benchmark domination, Thinking Machines engineered Inkling for adaptability and efficiency in real-world workflows. The standout feature of this release is Inkling's "controllable thinking effort." Developers can programmatically adjust the model's reasoning budget -- scaling from 0.2 to 0.99 -- to dictate how hard the AI should "think" before generating an output. As the company noted, "Inkling's continuous thinking effort lets you pick your point on the cost/performance curve -- reaching the same score with a fraction of the tokens". In practical terms, this allows enterprises to deploy Inkling with lower token expenditure for simpler tasks, while cranking up the compute overhead for complex, multi-step reasoning challenges. However, by keeping the thinking effort lower and generating fewer tokens, the cost-conscious enterprise can achieve high quality results and performance on simple tasks while spending less money, or, in the case of those running models locally, less costs on energy and compute resources. During the model's large-scale reinforcement learning (RL) training over 30 million rollouts, researchers observed an emergent phenomenon they called "chain of thought condensation". Over time, Inkling naturally learned to compress its internal reasoning steps -- dropping grammatical overhead and connectives -- while reaching the same accurate conclusions, resulting in drastically reduced latency. Epistemics and Censorship Resistance A notable element of Thinking Machines' release is its explicit focus on the model's epistemics -- specifically its calibration, instruction following, and resistance to censorship. In an ecosystem where open-weight models adopt either overly restrictive safety guardrails or echo state-aligned ideological talking points, Inkling was intentionally trained to answer directly on politically sensitive or heavily censored topics. To validate this approach, Thinking Machines submitted Inkling to the Propaganda and Censorship Eval developed by AI startup Cognition. According to the published findings, Inkling demonstrated "strong patterns of censorship non-compliance," effectively resisting ideological capture or boilerplate refusals when presented with sensitive subjects. Despite its resistance to censorship, the model maintains a robust defense against genuinely malicious, dangerous, or illegal queries. On the StrongREJECT benchmark -- which tests responses to unambiguous harmful requests -- Inkling scored 98.6%, placing it in line with strict frontier safety standards. Furthermore, on the FORTRESS benchmark, Inkling successfully navigated the line between safety and over-refusal: it achieved a 78.0% refusal rate on adversarial queries (such as those involving weapons, cyberattacks, or violence) while maintaining a 95.9% compliance rate on benign, look-alike queries. Thinking Machines noted that typical open-weight vulnerabilities remain within the architecture. Internal safety evaluations revealed an "occasional tendency to comply with role-play and indirectly framed prompts concerning harmful topics". The company advised enterprise developers to treat the model's built-in refusals as just one layer of security, recommending the downstream deployment of external moderation tools -- such as Llama Guard -- to filter adversarial jailbreaks and enforce use-case-specific safety policies at the application level. Under the Hood: Architecture and Multimodality Inkling's scale is staggering, yet sparse. The MoE architecture features 975 billion total parameters, but only 41 billion parameters are active during any given token generation. It supports a massive context window of 1 million tokens and diverges from typical transformer models by using relative positional embeddings instead of the industry-standard Rotary Positional Embedding (RoPE). True to the company's foundational vision, Inkling was trained from scratch to be natively multimodal. Unlike models that rely on bolted-on external encoders, Inkling uses an encoder-free early fusion approach. It directly ingests audio as discrete dMel spectrograms and visual data as 40x40 pixel patches via a hierarchical multi-layer perceptron (hMLP), projecting all modalities into a shared hidden space. Licensing: True Open-Source for the Enterprise For enterprise IT teams and developers, the most disruptive aspect of Inkling may be its licensing. Inkling is released under the permissive Apache 2.0 license. In an ecosystem where many so-called "open" models from Western labs are tethered to dual-use commercial licenses, acceptable use restrictions, or revenue caps, an Apache 2.0 designation makes Inkling a true open-source foundation. This gives developers the legal freedom to download, modify, integrate, and commercialize the model weights entirely royalty-free. The model is readily deployable across major open-source inference libraries -- including SGLang, vLLM, TokenSpeed, and llama.cpp -- and comes with a native NVFP4 quantized checkpoint optimized for NVIDIA Blackwell systems. Community Reactions: The Engineering Feat The AI community's response has been swift, praising both the model's openness and the underlying engineering execution. In a post on X, Thinking Machines co-founder John Schulman reflected on the rapid development cycle: "Inkling is out today, with open weights and in Tinker. It's been fun to watch this one come together: pretraining began last winter, and starting in mid-January a small team built up the coding, reasoning, and agentic training from there. We learned a lot building it, and I hope people find good uses for it." Horace He, a researcher at Thinking Machines (previously from PyTorch), underscored the difficulty of the task in another post on X: "It truly takes a village to release a model, perhaps especially an open weights model. Actually doing the entire process from scratch, from data to pretraining to posttraining to actual release, gives a lot of appreciation for anyone who does it!" The broader open-source ecosystem has also embraced the technical integrations. Lysandre Debut, the Chief Open-Source Officer at Hugging Face, shared his enthusiasm regarding the model's optimization in his own X post: "One thing I find quite striking is how much easier accelerating models has become... We replaced the model's causal Conv1D with the 'causal-conv1d' kernel. One line changed, +4% tokens per second. We then replaced its attention implementation with FlashAttention-4. Another single change, another +11%. That's a total throughput improvement of about 15%, without changing the model architecture or retraining anything." Tiezhen Wang, an ecosystem growth expert and ex-Googler, celebrated the release as a massive win for the open-source community, listing the model's impressive specifications on X, highlighting its "975B total, 41B active" size, "Native MTP support," and the highly coveted "Apache 2.0 license." Background: The Road to Inkling To understand the significance of Inkling, one has to look back at the rapid trajectory of Thinking Machines over the past 18 months. When Mira Murati departed OpenAI in late 2024 to found Thinking Machines alongside industry veterans like John Schulman and Barret Zoph, the stated goal was to pivot away from building isolated autonomous agents. Instead, the company aimed to build flexible, multimodal systems designed for genuine human-AI collaboration and open science. By July 2025, the startup had secured a historic $2 billion seed round led by Andreessen Horowitz at a $12 billion valuation. At the time, Murati promised the impending release of a product with a "significant open source component" to empower researchers and startups. The company's philosophy began coming into sharper focus in October 2025 with the launch of Tinker, a Python-based API for large language model fine-tuning that gave researchers granular control over training pipelines without the friction of distributed compute management. That same month, Thinking Machines researcher Rafael Rafailov delivered a provocative critique of the AI industry at TED AI. He argued that the current trajectory of simply throwing more compute at models was fundamentally flawed, noting that today's systems take shortcuts -- like wrapping code in blocks -- because they are trained strictly for task completion rather than genuine learning. Rafailov posited that the first artificial superintelligence would not be a "god model," but rather a "superhuman learner" capable of meta-learning and internalizing abstractions. Inkling's architecture -- specifically its controllable thinking effort and its ability to organically compress its chain of thought during RL -- feels like the first tangible realization of Rafailov's thesis. In May 2026, the lab teased its technical prowess with the research preview of TML-Interaction-Small, a system that eliminated "turn-based" chat by processing inputs and outputs simultaneously in 200ms chunks. This "full-duplex" breakthrough proved the company could build highly responsive, natively multimodal models from scratch. Now, with Inkling out in the wild, Thinking Machines has delivered on its foundational promises. By offering a massive, natively multimodal model under a true open-source license, they aren't just giving developers a new tool -- they are attempting to fundamentally rewrite the economics and accessibility of frontier AI development.

Murati's Thinking Machines releases first AI model for broad use | Fortune

Thinking Machines Lab is releasing its first AI model, more than a year after former OpenAI executive Mira Murati founded the artificial intelligence startup. The new model, called Inkling, is designed to be versatile and efficient, processing queries across different media while balancing "cost against performance," the company said on Wednesday. Inkling is also open-weight, meaning developers can download and customize the model without seeing the training data or source code. Thinking Machines said Inkling performs well on various benchmarks compared to similar open-weight offerings. The company said Inkling was trained "for strong performance across the board," but that "it is not the strongest model available today, closed or open." Founded last February by Murati and a cohort of other former OpenAI employees, Thinking Machines is part of a group of so-called AI neolabs attempting to build more advanced artificial intelligence software that can compete with leading labs like OpenAI and Anthropic PBC. Some of these companies, including Thinking Machines, have nabbed substantial sums from investors before releasing any products. Murati's startup raised $2 billion at a $12 billion valuation last year, and was said to have been in discussions about a bigger, subsequent round. Several employees have since left her company, joining firms like Meta Platforms Inc. and OpenAI. Read more: OpenAI, Meta, SpaceXAI Compete for More Cost-Efficient AI Models With its new model, Thinking Machines could be filling a void in the US market, which is thought to be lagging behind Chinese developers in putting out competitive open AI products. Meta, once a leader in open model development, has started to focus more on closed or proprietary models that customers would pay to use. OpenAI, meanwhile, released a pair of open models last year but its product portfolio remains dominated by paid offerings. As businesses and individuals become more conscious of their surging AI spending, some are turning to freely available models from China for certain tasks, raising national concerns in the US. Thinking Machines is not aiming to monetize the new model. Currently, the startup generates revenue through its developer tool called Tinker for fine-tuning -- or customizing -- AI models, which it sells to customers such as hedge fund Bridgewater Associates to improve AI performance on financial tasks. Inkling is also part of Murati's bid to build "interaction models" designed to let people collaborate more naturally with AI. "Our interactions with each other are very rich," she told Bloomberg last month. "There's a lot of information in our interactions - when we're silent, when we're thinking, when we're interrupting one another. Interaction models are able to capture all of these nuances."

Mira Murati's Thinking Machines drops Inkling, an open-weights model anyone can access

Mira Murati's Thinking Machines drops Inkling, an open-weights model anyone can access Mira Murati's Thinking Machines Lab Inc. today launched its first foundation model with the release of Inkling, making its full open weights available to developers so they can fine-tune it as they wish. Inkling is the first fully trained from scratch model released by Thinking Machines, coming after a year in which the company mostly made headlines for its sizable funding rounds and its partnership with Nvidia Corp. In a blog post, Thinking Machines explained that Inkling is a mixture-of-experts model that features 975 billion parameters, although for the average prompt it will only draw on a fraction of that number - around 41 billion - in order to process tasks faster and keep costs low. The company said the model was trained on about 45 trillion tokens of text, image, audio and video and is able to reason natively across all four inputs. However, its outputs are limited to text only, though that includes code, styled artifacts and structured data. The launch of Inkling suggests that Thinking Machines wants to provide the growing number of Western companies embracing lower-cost Chinese AI models with an alternative to those systems. That's because the model seemingly fills a gap in the Western open-source AI ecosystem, which has lagged far behind that of China's. That gap has only increased since Meta Platforms Inc. abandoned its Lama family of models in favor of a more proprietary approach with its latest AI systems. Murati, who was previously the Chief Technology Officer of OpenAI Group PBC before leaving in September 2024, has long insisted that her new company is all about accessibility, customization and multimodal collaboration, and that's clearly apparent with Inkling. Because Inkling is available to download with its full open weights, it means developers can look at the full codebase of the model, and tweak it for different use cases, without having to pay expensive licensing fees. It also features "thinking effort" controls that allow developers to make tradeoffs, such as sacrificing processing speed for accuracy. Uniquely, the model will also flag its outputs for uncertainty, instead of simply pushing out hallucinations. Developers can fine-tune the model directly on Tinker, which is the company's training application programming interface that launched in October. In its early test results, Thinking Machines showed that Inkling was able to achieve a comparable coding performance with Nvidia's Nemotron 3 Ultra model, despite using two-thirds less tokens. Futurum Group analyst Mitch Ashely told the Wall Street Journal that the open-weight model ecosystem has been dominated by Chinese AI firms for the last year, and that Inkling is the first Western alternative to those systems. "It gives Western enterprises a credible alternative positioned on customization economics, shifting spend from per-token API pricing to infrastructure the enterprise controls," he said. "Engineering teams should treat base-model selection as an architecture decision. The model an organization fine-tunes becomes part of its software substrate and switching costs compound with every downstream customization. That evaluation cannot be deferred." Thinking Machines acknowledged that Inkling isn't as strong as some of the most advanced proprietary AI systems available, but it's clearly betting that its customizability will make up for that. Instead of making Inkling available in a rigid chatbot-style app, it's positioned as a base model that organizations should fine-tune and run themselves on their own infrastructure. It's a strategy that should appeal to many organizations. In a collaboration with Bridgewater Associates LP, researchers used the Tinker platform to fine-tune an open model with specialized financial data. They ended up with a low-cost, lightweight model that scored an impressive 84.7% on leading financial reasoning benchmarks, outperforming the most advanced proprietary alternatives at less than 10% of the cost. Thinking Machines said it was able to develop Inkling from scratch in less than nine months, which is only a fraction of the multiyear development timelines seen at rivals like OpenAI and Anthropic PBC. In its post on X announcing the model, the company explained it was trained on Nvidia's GB300 NVL72 system under a partnership between the two firms that was announced in March. Instead of charging customers for access via a metered API, Thinking Machines plans to generate revenue through Tinker, which is a paid service that makes it simple for developers to fine-tune open weights models for specific tasks. It should be a key test that will show whether or not open-weight AI models have what it takes to disrupt the gated, paid access model pioneered by Silicon Valley's biggest AI firms.

Mira Murati's AI start-up unveils customisable, open source model Inkling

Inkling is Thinking Machines Labs' first big AI model launch, and comes a year after the start-up was founded. Mira Murati's Thinking Machines Lab has unveiled the first of its family of multimodal open source models called Inkling. The model was the first from the start-up to be trained on the Nvidia GB300 NVL72 systems, following a partnership between the two companies earlier this year. Inkling does not promise to be the strongest open or closed model available, but rather markets itself as one made to be customised. It has 975bn total parameters, with 41bn of active ones and supports a context window of up to 1m tokens. "Inkling is designed to be broad. We trained it across agentic, reasoning, coding, instruction-following, factuality, vision, and audio tasks, rather than narrowly optimizing for one domain," Thinking Machines Lab said in a blog post announcing the new model on Wednesday (15 July). The model, available to download on Hugging Face, was trained using 45trn tokens of text, images, audio and video. It can input text, images and audio and output only text. It is a Mixture-of-Experts transformer that follows a similar design to DeepSeek-V3. Inkling's model card places it squarely around its open-weight contemporaries, including Kimi K2.5, Kimi K2.6, GLM 5.2 and DeepSeekV4 Pro, across benchmarks. The model performs much better producing audio outputs than comparable open weights and closed weights models. The launch represents Thinking Machines Labs' first major AI model showcase after more than a year under development. Murati, who left OpenAI as its chief technology officer in 2024, founded her start-up just months after, with plans to make "AI systems more widely understood, customisable and generally capable". The start-up quickly raised $2bn in July last year to hit a $12bn valuation, with reports from November suggesting that Thinking Machines Lab was already readying to raise funds at a valuation of $50bn. Alongside Inkling, the company is also launching Inkling-Small in preview - a lighter-weight model with 12bn active parameters that supposedly achieves strong performance with "even lower cost and latency". Nvidia has taken a liking to Thinking Machines Lab, which made a "significant" investment into the start-up alongside taking the company on for a multi-year partnership to develop its AI models. The partnership is expected to have cost the chipmaker several billions. Don't miss out on the knowledge you need to succeed. Sign up for the Daily Brief, Silicon Republic's digest of need-to-know sci-tech news.

OpenAI's Former CTO Just Dropped a Very Different Kind of AI Model

Inkling was released yesterday, and on its webpage Thinking Machines promises its "mission is to build AI that extends human will and judgment." More specifically, it calls Inkling "a model we trained from scratch with the full weights available, so that people can make it their own." Here's what "full weights available" actually means: For generative AI models like Inkling, ChatGPT, and many rivals, weights are essentially a vast array of numbers that constitute the training data set that makes the AI magic work. If you train a model with particular words over and over, the "weights" in the model linked to these words get tweaked, which influences the answers the AI gives to questions later on. For many mainstream AI models, these weights are proprietary and secret. When you query such models, for a vast variety of purposes -- from writing code to drumming up a quick blurb for your website -- you have to blindly trust that they're working the way you want them to, and hope that the output is a good fit for your needs. This is where Inkling stands out. Through a special tool called Tinker, users can shape and customize how Inkling works so that it aligns more precisely with their individual needs -- which means you can tweak the model's weights so that its outputs are a better fit to what you want. Thinking Machines explains that Inkling is "designed to be broad," and has been trained "across agentic, reasoning, coding, instruction-following, factuality, vision, and audio tasks, rather than narrowly optimizing for one domain." This makes it ideal for "customization and real-world use: different users need models that can adapt to very different workflows," the company says, arguing that this idea is better than an AI excelling on mere benchmark tests.

Mira Murati's AI start-up unveils customisable model Inkling

Inkling is Thinking Machines Lab's first big AI model launch, and comes a little more than a year after the start-up was founded. Mira Murati's Thinking Machines Lab has unveiled the first of its family of multimodal open-source models called Inkling. The model was the first from the start-up to be trained on Nvidia GB300 NVL72 systems, following a partnership agreement between the two companies earlier this year. Inkling does not promise to be the strongest open or closed model available, but rather markets itself as one made to be customised. It has 975bn total parameters, with 41bn active, and supports a context window of up to 1m tokens. "Inkling is designed to be broad. We trained it across agentic, reasoning, coding, instruction-following, factuality, vision, and audio tasks, rather than narrowly optimising for one domain," Thinking Machines Lab said in a blogpost announcing the new model on Wednesday (15 July). The model, available to download on Hugging Face, was trained using 45trn tokens of text, images, audio and video. It can input text, images and audio and output only text. It is a Mixture-of-Experts transformer that follows a similar design to DeepSeek-V3. Inkling's model card places it squarely around its open-weight contemporaries, including Kimi K2.5, Kimi K2.6, GLM 5.2 and DeepSeekV4 Pro, across benchmarks. The model performs much better producing audio outputs than comparable open-weight and closed-weight models. The launch represents Thinking Machines Lab's first major AI model showcase after more than a year under development. Murati, who left OpenAI as its chief technology officer in 2024, founded her start-up just months after, with plans to make "AI systems more widely understood, customisable and generally capable". The start-up quickly raised $2bn in July last year to hit a $12bn valuation, with reports from November suggesting that Thinking Machines Lab was already readying to raise funds at a valuation of $50bn. Alongside Inkling, the company is also launching, in preview, Inkling-Small - a lighter-weight model with 12bn active parameters that supposedly achieves strong performance with "even lower cost and latency". Nvidia has taken a liking to Thinking Machines Lab, having made a "significant" investment into the start-up alongside taking the company on for a multi-year partnership to develop its AI models. The partnership is expected to have cost the chipmaker several billion dollars. Don't miss out on the knowledge you need to succeed. Sign up for the Daily Brief, Silicon Republic's digest of need-to-know sci-tech news.

Thinking Machines: AI startup Thinking Machines launches an open-weight AI model

Named Inkling, the model is open-weight, meaning users can download, run and customize the underlying systems, unlike proprietary, ⁠closed-source ⁠models. AI startup Thinking Machines revealed on Wednesday a new artificial intelligence model that could serve as one of the few alternatives to popular open-source offerings from Chinese AI labs. Named Inkling, the model is open-weight, meaning users can download, run and customize the underlying systems, unlike proprietary, ⁠closed-source ⁠models. It is the first general-purpose model release to come out of Thinking Machines, a San Francisco-based startup founded last year by OpenAI's former chief technology officer Mira Murati. Thinking Machines launched its first product called Tinker, which helps customize AI models, last October. Inkling is available on Tinker and other developer platforms, it said. The ⁠model has 975 billion parameters - variables that determine how an AI system processes information - making it one of the largest ⁠models of its kind. The open-source ecosystem in the West lags behind its counterpart in China, especially in the wake of a void left by Meta, which changed course to a proprietary approach after the disappointing release of its open Llama 4 model last year. Businesses have in turn flocked to adopt Chinese models as the primary alternatives to expensive closed-source models. Hedge fund Bridgewater Associates used Tinker to build a custom version of Qwen, a model developed by China's Alibaba, which it said ⁠outperformed top proprietary models at lower costs. Thinking Machines published a series of benchmarks that compared Inkling's capabilities with closed models from Anthropic, Google, and OpenAI, as well as leading open offerings, most of them from Chinese labs. While those other models maintain the edge on performance overall, Inkling put in a competitive showing, particularly on agent-related tasks, that could spur interest from prospective users.

Mira Murati Unveils Open-Weight AI Model as Alex Karp, Satya Nadella Warn Companies Are Giving Away Their

Her startup, Thinking Machines Lab, recently released "Inkling," an open-weight AI model designed to let businesses customize AI without exposing their proprietary data. The Enterprise Data Crisis The launch of Inkling arrives exactly as major tech executives warn that relying on frontier AI labs is jeopardizing enterprise data and competitive advantages. Decentralized Control and the 'Inkling' Solution Murati's new release directly addresses these enterprise fears. Inkling is an open-weight model, meaning developers can download and tailor the software to their needs without ever seeing the training data or source code. Venture capitalist Bill Gurley highlighted the strategic timing of the launch. On X, Gurley noted that Murati's approach is "quite consistent with the POV this past week from Alex Karp & Satya about companies controlling their own IP," calling the launch "Right place. Right time. Decentralized." Murati herself emphasized the necessity of decentralized AI in a recently published manifesto. "Human values don't average out," Murati stated. "Local knowledge can't be centralized." Her company argues that AI should help organizations cultivate their unique knowledge, rather than extract a snapshot of it to create a standard, centralized offering. Here's a list of a few AI-linked exchange-traded funds, spanning broad AI and technology, robotics and automation, active and generative AI, and AI infrastructure and energy. Disclaimer: This content was partially produced with the help of AI tools and was reviewed and published by Benzinga editors. Photo courtesy: Shutterstock Market News and Data brought to you by Benzinga APIs To add Benzinga News as your preferred source on Google, click here.

Thinking Machines Releases First In-House AI Model, And It is Open Source

The timing couldn't be better for ex-OpenAI CTO Mira Murati's Thinking Machines Lab to release their first AI model. Called Inkling, it is unlike other flagship models from AI labs as it is open-weight and outside developers and companies can download and modify it directly. "Our model, called Inkling, is a Mixture-of-Experts transformer with 975B total parameters, 41B active. It supports a context window of up to 1M tokens. It was pretrained on 45 trillion tokens of text, images, audio and video," the company says in a blog post while noting that it is the first among a family of models of different sizes. Thinking Machines Labs also announced that they're sharing "a preview of Inkling-Small, a lighter-weight model with 12B active parameters, trained with a similar recipe, that achieves strong performance with even lower cost and latency." When we said the timing is just about perfect, we were referring to the recent spate of comments from tech czars in the United States questioning closed models of the kind that OpenAI, Anthropic and Google have been launching with ChatGPT, Claude, and Gemini, respectively. These are built as general-purpose chatbots first with agentic and autonomous features layered on top. In fact, Microsoft CEO Satya Nadella, whose company invested the early billions into OpenAI and Anthropic, warned the world about the pitfalls of enterprises using such proprietary AI models and effectively paying twicefor them - subscription costs and giving away their business knowledge embedded in prompts and corrections. Even Thinking Machines had published a blog post last week that now appears to have set the background for last night's launch. It had claimed that AI trained by one company and then hardcoded starts to underperform on the AI front as enterprises shape-shift over time. This happens because those operating the company have expertise that's unique to them and no hardcoded AI training can capture it on the fly. The blog highlighted the values of "decentralised alignment" arguing that even with the best of intentions, a model shaped at one place inevitably encodes the values of its owner, not the individual users it serves. Each AI lab trains its next flagship models using its previous edition so whatever character emerges from that loop, everyone gets the same, each generation inheriting the traits of the last. Therefore, organisations and individuals must align AI to their own values that must then get encoded in model weights, the blog said. "If the user's values and desires only impact the model through a prompt, the user finds that surface properties change while the deeper habits remain. Allowing core model behaviour to change significantly with prompts sacrifices safety, making a malleable centralized model vulnerable to repeated attacks," the blog said. Inkling has been designed as a mixture-of-experts system with over 975 billion parameters but draws from only 41 billion parameters for any task, thus making it a design that keeps very large models faster and cheaper to run. Additionally, it was trained on 45 trillion tokens of text, image, audio and video as well as reasons natively across all four, the company's post says. For now, Thinking Machines Labs says the outputs would be limited to text, including code, styled artefacts, and structured data. This is the first public proof from the Labs after more than 18 months building AI infrastructure, mostly under the radar. The company had released some parts of their work via a research preview of interaction models or AI that is designed to listen and speak as well as interrupt instead of waiting as users do with chatbots. Inkling is structurally designed to provide calibrated answers that includes flagging uncertainty (does this mean less of hallucinating?). It also uses a third as many tokens as Nvidia's Nemotrol 3 Ultra (the latest generation open-weight model) in order to hit the same coding performance. Of course, none of these make Inkling superior to others in the market. And the company admits this openly in the blog post which states that Inkling is "not the strongest overall model available today, open or closed." Given the fact, the next obvious question would be which segment of the enterprise market is Thinking Machines targeting? The company says it proposes to market Inkling less as a finished product than as a starting point in an AI-led journey for enterprises to fine tune themselves using Tinker, which is Thinking Machines' model-customisation platform. So, it is up to the customers to ensure safety while customising the model. In fact, the company shared a recent use case with Bridgewater Associates (one of the largest hedge funds) whereby researchers took an existing open-source model and trained it further on the latter's financial expertise. The outcome scored 84.7% on financial reasoning tests, which was higher than proprietary AI models. What's more, it cost them one-fourteenth the cost of existing models to operate. The company did a bit of gloating in the blog by stating that Thinking Machines took just about nine months to get to this point compared to about five years that OpenAI took to bring the technology to the market and generate revenues. Anthropic did the same but in two years lesser timeframe than OpenAI. Another bit of information that the blog shared related to distillation with Thinking Machines stating that this was partly true with Inkling. They pre-trained the model from scratch but used open-weight models such as Moonshot AI's Kimi K2.5 to generate some early post-training data. However, the next model would be totally indigenous, the company adds. One aspect where Thinking Machines has played safe relates to its partnership with Nvidia where it signed up to deploy a gigawatt of Vera Rubin compute capacity while training Inkling totally on the Nvidia GB300 NVL72 systems. The company is silent on how it would cover the costs or the actual revenue model it intends to follow. Of course, this brings up the question around whether Mira Murati and her team would ever reach the spending levels scaled by OpenAI or Anthropic. Or does the efficiency-first approach mean that the overall economics of the company would look totally different from what we have experienced over the past several months. Would it spend as much as its larger rivals or rest assured that it won't need to at all. The company said Inkling became available on Tinker immediately with context length options of 64K and 256K tokens. Additionally, the company offered a 50% discount for a limited time, with full pricing information available here for prospective users. It is also available via APIs on TogetherAI, Fireworks, Modal, Databricks, and Baseten. "We worked with RadixArk to provide open-source inference and RL support in SGLang and Miles. We worked with Inferact to support inference in vLLM, with Lightseek for inference in TokenSpeed, and with Unsloth for inference in llama.cpp. Finally, we partnered with Hugging Face on integration with transformers," the blog post said.