Mira Murati’s Inkling AI model arrived this week, and it shifts who gets to shape a frontier system. Thinking Machines Lab posted the full weights on Hugging Face. You can download them, run the model on your own hardware, and rebuild it for your work. That reach sets it apart from the closed flagships sold by OpenAI, Anthropic, and Google.
What Inkling is
Inkling is a Mixture-of-Experts model with 975 billion total parameters. It calls on about 41 billion for any single task, which keeps a very large system faster and cheaper to run. The model handles a context window of up to 1 million tokens. It reads text, images, and audio, and reasons across all three. Thinking Machines trained it on 45 trillion tokens of text, images, audio, and video.
The company is blunt about where the model lands. In its own words, Inkling is not the strongest model available today, open or closed. The pitch runs a different way. Murati’s team wants you to own a base you can adapt, not rent a locked system you cannot see inside.
Why Mira Murati’s Inkling AI model targets builders
Mira Murati’s Inkling AI model is built to be changed. Thinking Machines made it available for customization on the Tinker fine-tuning platform on launch day. You point the model at your own data, train it for your domain, and keep the result. The company also added an Inkling Playground in the Tinker console, so you can chat with the model before committing to a training run.
Controllable thinking effort is the other lever. You can dial how long the model reasons, trading speed for depth depending on the task. A lighter preview, Inkling-Small, runs 12 billion active parameters and aims for strong results at lower cost and latency.
Who Murati is and why it counts
Murati left OpenAI as chief technology officer in 2024. Months later, she founded Thinking Machines Lab, which raised a $2 billion seed round at a $12 billion valuation before it shipped a single product. Inkling is the lab’s first in-house model, built in under a year and trained on NVIDIA GB300 NVL72 systems.
Her open-weight AI model choice carries a history. At OpenAI in 2019, the lab held back the full GPT-2 over misuse fears. Murati now signals a case-by-case path: release openly when the risk looks manageable, hold back when it does not. Open weights this time do not promise open weights next time.
What it means for you
The stakes reach past the research world. If you run a business, an open-weight AI model you can fine-tune keeps your proprietary knowledge in-house instead of feeding it through someone else’s API. Thinking Machines points to work with hedge fund Bridgewater, where a fine-tuned open model scored 84.7 percent on a financial reasoning test. That figure came from the two companies’ own evaluation, not an independent one, so weigh it as a first-party claim until others confirm it.
For developers, Mira Murati’s Inkling AI model widens the menu. You get a large, multimodal base you can inspect, adapt, and deploy through providers like TogetherAI, Fireworks, and Baseten. The Mixture-of-Experts model design keeps running costs down while the parameter count stays high.
Inkling is the first in a planned family from Thinking Machines Lab. More models will follow. Whether each future model ships with open weights is not settled yet.
