Thinking Machines Lab, the AI startup founded by former OpenAI chief technology officer Mira Murati, has released its first model, an open-weight system called Inkling. It is a genuine milestone for one of the most closely watched and heavily funded new labs in the field, one that raised roughly 2 billion dollars at a 12 billion dollar valuation before shipping a single product, and the model is a direct expression of the company's central bet, that AI which organizations can adapt for themselves will beat the one-size-fits-all models the largest labs currently sell.
Inkling is a mixture-of-experts model with 975 billion total parameters, of which about 41 billion are active for any given task, a design that keeps the compute cost down while giving the model a very large base of knowledge to draw on. It was trained on 45 trillion tokens spanning text, image, audio, and video, so it reasons across all of those inputs natively rather than bolting extra senses onto a text model. Two of its choices stand out, it gives calibrated answers that flag their own uncertainty instead of confidently guessing, and it lets a user turn thinking effort up or down to trade speed for quality.
What is most striking is the framing. Thinking Machines says plainly that Inkling is not the strongest model available today, closed or open. Rather than chase the top of a leaderboard, it is aiming for something well-rounded, efficient, and above all adaptable. On coding tasks it reportedly uses roughly a third of the tokens a comparable model needs for the same result, and in a case study with the investment firm Bridgewater it reached 84.7 percent accuracy on financial reasoning at a small fraction of the cost. The pitch is not to be the best general chatbot, it is to be a strong base that others can shape.
That is where the lab's larger plan comes into focus. Inkling is meant to be customized through Tinker, Thinking Machines' fine-tuning platform, so a company can adapt it to its own data and workflows instead of renting a frozen model whose behavior is controlled elsewhere. It is the concrete version of a thesis Murati and her team laid out this month, that the future of AI is distributed and customizable, shaped by the people using it, rather than a handful of giant closed models that everyone rents on the same terms.
Why it matters is that Inkling turns a contrarian argument into something you can actually download. The dominant strategy in AI has been to build ever larger frontier models and sell access to them, and most of the attention still goes to whoever posts the highest benchmark scores. Thinking Machines is wagering that openness and adaptability, an open-weight model plus the tools to tune it, will matter more to a lot of organizations than raw supremacy on a test. Whether that bet pays off is unproven, and by the company's own admission this is not the most powerful model on the market. But coming from Murati's team, with the money and the spotlight it commands, it is one of the most credible challenges yet to the idea that the winners in AI will simply be whoever trains the single biggest brain.