Mira Murati launches 975B-parameter answer to Chinese open AI

Signals Inbox·July 15, 2026·Frontier AI

Mira Murati's Thinking Machines Lab has released Inkling, a 975-billion-parameter model with fully open weights. The headline number is huge, but the real tension is geopolitical: Chinese labs have been pulling ahead in open AI, and one of America's most heavily funded new labs is finally shipping an answer.

Q1What actually launched?

Thinking Machines Lab officially released Inkling, its first model trained from scratch. It is a general-purpose model that can understand text, images and audio, generate text, reason, write code and be fine-tuned for specific jobs. Its weights are available under an Apache 2.0 license, so developers can download, host and modify it.

Q2How big is 975 billion parameters?

Very big, but slightly misleading on its own. Inkling uses a Mixture-of-Experts design, so only about 41 billion of its 975 billion parameters are active for each token. That makes inference much lighter than running the whole model at once. It was trained on roughly 45 trillion tokens and supports up to one million tokens of context when self-hosted.

Q3Is it actually America's answer to Chinese open AI?

It is an answer, but not a clean victory. Independent testing places Inkling ahead of other major US open-weight releases, including Nemotron 3 Ultra. But Thinking Machines openly says it is not the strongest model overall, and Chinese models such as GLM 5.2 still beat it on several tests. The important change is that America now has another serious, large and customizable open model in the race.

Q4Why release the weights instead of another chatbot?

Because Thinking Machines is betting that companies want control. A closed chatbot gives everyone roughly the same model through an API. Open weights let a hospital, bank or software company run Inkling on its own infrastructure and train it around private data or expert workflows. Tinker removes some of the painful infrastructure work by letting developers fine-tune the model through an API.

Q5Does the huge size make Inkling the best model?

No. Parameter count is not a scoreboard. Training data, architecture, post-training and inference design can matter more. Inkling performs well in coding and reasoning, but it remains behind leading closed models from OpenAI and Anthropic and behind some Chinese open models. Thinking Machines is selling it as a flexible base to customize, not as the automatic winner of every benchmark.

Q6So why does this matter now?

Because Thinking Machines raised $2 billion at a $12 billion valuation before proving it could build a major model. Inkling is the first real test of that bet. It also arrives as Chinese open models gain users because they are capable, cheap and easy to modify. The next question is not whether Inkling looks impressive on launch day. It is whether developers fine-tune it, deploy it and choose it over stronger Chinese alternatives.