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

Thinking Machines Lab, the startup Mira Murati built after leaving OpenAI as CTO, put out its first in-house model Wednesday morning. It's called Inkling, and the headline detail is that it's open-weight — anyone can download it and start tinkering, which is not how OpenAI, Anthropic, or Google ship their flagship products. Under the hood it's a mixture-of-experts system with 975 billion total parameters, though only about 41 billion get used for any given task, a common trick for keeping huge models affordable to run. It trained on 45 trillion tokens spanning text, image, audio, and video, and the company says it reasons across all four natively — but for now, the outputs coming out the other end are limited to text, code, and structured data.

This is the company's first real public proof point after roughly a year and a half of quiet infrastructure work, following a May preview of "interaction models" designed to listen and interrupt rather than just wait their turn. Inkling is meant to give calibrated answers, flagging uncertainty instead of bluffing, and it lets users crank a "thinking effort" dial up or down depending on whether they want speed or quality. On one benchmark, the company claims Inkling matches Nvidia's Nemotron 3 Ultra on coding while using a third as many tokens. And yet Thinking Machines is upfront that Inkling isn't the strongest model out there, open or closed — it's chasing well-rounded rather than best-in-class.

That framing only makes sense once you see how the company wants Inkling used. It's positioning the model less as a finished product and more as raw material, something organizations fine-tune themselves through Tinker, its customization platform — which also means customers, not Thinking Machines, own the safety implications of whatever they build on top. That's a sharp departure from how OpenAI, Anthropic, and Google operate, all of them built around general-purpose chatbots with agentic features bolted on afterward.

The company has been laying philosophical groundwork for this. A post last week argued that centrally trained, frozen models underperform ones organizations can shape around their own expertise. Microsoft's Satya Nadella made a related point in a Sunday blog post, warning that companies leaning on proprietary AI effectively pay twice — once in subscription fees, again by feeding proprietary knowledge back into future model versions. Hugging Face's Clem Delangue predicted something similar: frontier models handling experimentation, private or open alternatives doing the actual production work. Thinking Machines' own case study, a joint project with hedge fund Bridgewater Associates, reported an 84.7% score on financial reasoning tests using a fine-tuned open model, beating proprietary competitors at roughly a fourteenth of the running cost — though that number comes from the two companies' own testing, not an outside evaluator.

What's striking is the pace. OpenAI took roughly five years to reach market with revenue, Anthropic about three; Thinking Machines says it did this in nine months. It's also partly honest about how it got there — Inkling was pre-trained from scratch, but some early post-training data leaned on other open-weight models, including Moonshot AI's Kimi K2.5, before reinforcement learning took over. The company says its next model will skip that step entirely. Money remains the murkier part of the story: a gigawatt Nvidia partnership and training runs on GB300 NVL72 hardware suggest serious spend, but a reported $50 billion raise reportedly stalled earlier this year and the company hasn't discussed funding since. Because the weights are public, nobody who downloads Inkling owes Thinking Machines anything to run it — so the actual business has to come from Tinker, not the model itself.

Headcount, at least, sounds stabilized at roughly 200 people, following a stretch of departures this year that included two co-founders leaving for OpenAI in January. The company reportedly prizes continuity over star power internally — a notable stance for an outfit whose public identity is still so tied up in one very famous name.