Member-only story
Thinking Machines shipped it under the Apache 2.0 licence, the most permissive licence there is. It tops no leaderboard, because it was built to be fine-tuned into yours, not to win one.
On 15 July 2026, Mira Murati’s Thinking Machines Lab released its first model, Inkling, and then did something labs almost never do on launch day: it published the benchmarks it loses. On the model card, Inkling posts 77.6% on SWE-bench, verified against Claude Fable 5’s 95.0, 46.0% on Humanity’s Last Exam with tools against Fable 5’s 64.5, and 87.2% on GPQA Diamond against the 94.1 the closed frontier is running. It does not top any of the headline leaderboards. For a debut from the most-watched startup in AI, founded by the woman who ran research at OpenAI, that reads at first like a miss.
It is not a miss. It is the whole strategy, and the tell is the licence. Inkling is a 975-billion-parameter multimodal model released under Apache 2.0, which is to say you can download it, run it, fine-tune it, sell what you build on it, and strip it for parts, with no permission and no strings. That is the most permissive terms a frontier-scale model has ever shipped under. Murati’s answer to the…