Thinking Machines Lab Drops Its First Model

Thinking Machines Lab, the startup built by a cluster of OpenAI alumni, has finally shipped something you can actually touch. The model is called Inkling, and unlike the walled-off systems from the bigger labs, it's open-weight — meaning researchers and startups can pull it down and start modifying it themselves. The company built Inkling from the ground up to handle audio and video alongside text, and while it says the model doesn't top the popular benchmark charts, it holds its own across a range of tasks and can do serious reasoning and coding work. It's also enormous: 975 billion parameters, the kind of size that demands a cluster of specialized chips rather than a laptop.

One detail buried in the company's own blog post is more interesting than the spec sheet. Thinking Machines used Inkling to fine-tune and improve itself, a small but telling example of AI increasingly building AI. During that process, something unexpected happened to the model's reasoning trail. Inkling normally narrates its thinking in plain language, but as training pushed it toward efficiency, that chain of thought got noticeably tighter — shedding grammatical filler while staying understandable, and without changing the quality of its final answers.

For Thinking Machines, the release is as much about positioning as engineering. Open-weight models have caught on because they're cheaper to run and easier to bend toward specific tasks than the fee-gated closed alternatives. Right now, the strongest open-weight models come out of China, and Thinking Machines is making a direct claim here — it says Inkling performs at a similar level. Whether that holds up under scrutiny will matter a lot for a company still trying to prove it's more than a roster of famous names.

And the open release lines up neatly with the philosophy the company laid out in an earlier post: AI shouldn't sit in the hands of a small number of companies, and more people should be able to build their own models on their own data. That's a comfortable stance to take from a well-funded position, but it's also a real product decision, not just a talking point.

The founders' résumés explain some of the attention. Thinking Machines launched in February 2025 with Mira Murati, who was OpenAI's CTO and briefly its CEO, John Schulman, an OpenAI cofounder who helped build ChatGPT, and Lilian Weng, a former OpenAI VP who ran safety and robotics work. The startup raised the largest seed round ever recorded, arriving with a $12 billion valuation before shipping a single public product. Before Inkling, it had put out Tinker, a fine-tuning tool, demoed a natural voice interaction feature, and published research.

OpenAI may have set off the current AI rush with ChatGPT, but the defectors are carving out their own territory. Anthropic, another OpenAI-adjacent company, just filed for an IPO that could value it above a trillion dollars, with its Claude models winning over businesses for coding work in particular. Thinking Machines dropping an open-weight model into that mix is a statement that the defector wave isn't finished expanding.