Thinking Machines Lab, the AI startup founded by former OpenAI executives Mira Murati, John Schulman, and Lilian Weng, has released its first in-house model. Called Inkling, the open-weight system is a 975-billion-parameter mixture-of-experts model trained from scratch on text, images, audio, and video. It is the company's first major public proof point after more than a year of building largely out of sight.
The release is significant not because Inkling claims to dominate every benchmark. It doesn't. Thinking Machines explicitly says the model is "not the strongest overall model available today, open or closed." What matters is what Inkling represents: a bet that open, customizable AI will win in the enterprise market, and a test of whether a startup with 200 employees and a $12 billion valuation can compete with labs spending orders of magnitude more.
The Architecture Behind Inkling
Inkling uses a mixture-of-experts architecture with 975 billion total parameters, but it activates only about 41 billion for any given task. This is a well-established design pattern for keeping very large models practical. The model was trained on 45 trillion tokens across four modalities: text, image, audio, and video. It reasons natively across all four, though for now its outputs are limited to text, including code, styled artifacts, and structured data.
One of the more interesting details to emerge from the company's release materials involves how Inkling reasons. Like many large models, Inkling generates a chain-of-thought explanation for its complex reasoning. But the company observed something unusual during training: over time, Inkling's chain-of-thought became more concise, dropping grammatical overhead while remaining comprehensible. The model essentially learned to think more efficiently, a phenomenon that could have implications for how we train future generations of models.
Inkling also lets users dial its "thinking effort" up or down, trading depth for speed. On one internal benchmark, the company claims Inkling uses a third as many tokens as Nvidia's Nemotron 3 Ultra to hit the same coding performance.
An Open-Weight Bet Against the Giants
Thinking Machines is positioning Inkling as the starting point for organizations that want to customize their own AI, not as a finished product. The model is available through Tinker, the company's model-customization platform, which means customers take responsibility for fine-tuning and safety alignment. This is a fundamentally different approach from OpenAI, Anthropic, and Google, which sell ChatGPT, Claude, and Gemini as general-purpose products with agentic features layered on top.
The strategic logic is laid out in a blog post the company published last week. AI trained centrally by one company and frozen in time, Thinking Machines argues, underperforms AI that organizations tailor to their own expertise. The argument is gaining traction. Microsoft CEO Satya Nadella warned in a recent post that enterprises using proprietary AI models pay twice: once in subscription costs and again by contributing proprietary knowledge through prompts and corrections. Hugging Face CEO Clem Delangue has predicted that most production AI work will shift to private or open-source alternatives.
The clearest validation may have come from a project with Bridgewater Associates, the world's largest hedge fund. Researchers from both companies took an existing open-source model and fine-tuned it on Bridgewater's own financial expertise. The result scored 84.7% on financial reasoning tests, beating top proprietary models while costing roughly one-fourteenth as much to run. The results are self-reported, not independently verified, but they illustrate the thesis Thinking Machines is building around.
Revenue, Costs, and the Open-Source Economy
Thinking Machines has been tight-lipped about its finances. A reported $50 billion fundraising round was said to be coming together last November but stalled by January, and the company has declined to comment on its funding picture since. It trained Inkling entirely on Nvidia's GB300 NVL72 systems, part of a broader partnership with Nvidia to deploy a gigawatt of Vera Rubin computing capacity, but has not said how it plans to cover those costs.
The economic question at the heart of Thinking Machines' model is unusual. OpenAI and Anthropic generate revenue by metering access to their models. Thinking Machines cannot do the same because its weights are publicly downloadable. The company's revenue has to come from Tinker: training, fine-tuning, and a cut of the hosting ecosystem built around the model. It is a bet that the value is in the customization layer, not the raw model itself, and it is a bet that has no direct precedent at this scale.
Headcount, at least, appears stable. The company now employs roughly 200 people, recovering from a wave of departures earlier this year that included two co-founders who left for OpenAI in January. The company's culture, according to a source inside the organization, deliberately de-emphasizes individual personalities in favor of institutional continuity. That is a notable stance for a startup whose story is still so tightly associated with Mira Murati, whether she planned it that way or not.
What Happens Next
Inkling is a first step, not a final destination. The company says its next model will use fully self-contained post-training, eliminating the need to draw on outputs from other open-weight models for early training data. That matters because the industry is increasingly scrutinizing the practice of distillation, and Thinking Machines has acknowledged that Inkling's early post-training data partially relied on models like Moonshot AI's Kimi K2.5.
For founders and builders, Inkling represents something worth watching closely. If Thinking Machines can make the math work, it could validate a new model for AI companies: open-weight by default, revenue from customization, and competition with closed labs on speed of iteration rather than scale of spending. If it cannot, the lesson is equally valuable. Either way, the experiment is now running in public, and the results will tell us a great deal about where the AI industry is headed.