Mira Murati's Lab Ships Its First Open Model

Thinking Machines Lab spent most of its first year saying almost nothing. The startup, founded by former OpenAI chief technology officer Mira Murati and stocked with senior researchers from the same building, raised an enormous seed round and then went quiet. On July 15 it finally showed its hand, releasing Inkling, a 975-billion-parameter model that anyone can download and run.

The headline choice is the interesting part. Inkling ships as open weights, with the full model posted to Hugging Face. That puts Murati's lab on the opposite side of a line from her former employer, which keeps its frontier models locked behind an API.

What is under the hood

Inkling is a mixture-of-experts transformer, a design that holds a very large number of parameters but only fires a fraction of them for any given token. Of its 975 billion total parameters, 41 billion are active at a time, which keeps the running cost closer to a much smaller model. It handles a context window of up to one million tokens and was pretrained on 45 trillion tokens spanning text, images, audio and video.

It is natively multimodal rather than a text model with vision bolted on afterwards. It reads images as small pixel patches and takes audio in as spectrograms, and it carries a feature the lab calls controllable thinking effort, a dial from 0.2 to 0.99 that lets a developer trade speed for deliberation on a given task.

The claims, and the caveats

Thinking Machines is careful not to oversell. Its own write-up describes Inkling as "a good open-weights base for customization" rather than the strongest model available, which is a more honest framing than the field usually manages. On agentic coding it reports 77.6 percent on SWE-bench Verified, and on the FORTRESS adversarial safety benchmark it reports 78 percent. Those are respectable figures for an open model, and the lab presents them as a floor to fine-tune up from rather than a finished product. A smaller preview, Inkling-Small, ships alongside it at 276 billion parameters with 12 billion active.

Why open, why now

The release lands in a market that has quietly shifted under the frontier labs. American companies have been reaching for cheap Chinese open models when the closed options cost too much, and Meta has started charging for its own weights. An open model from a well-funded American lab with this pedigree changes the shape of that conversation. It gives companies wary of sending data to someone else's API a serious model they can host themselves, and it hands researchers something they can pull apart and study.

Whether it dents the leaders is a separate question. Open weights lower the cost of switching, but the frontier of raw capability still belongs to the closed labs for now. What Inkling does is make the middle of the market more crowded, and that usually turns out well for the people buying rather than the ones selling.