Here is a rare thing in AI: a smaller, cheaper model that beats the bigger one it was distilled from. On 2 August, Mira Murati’s Thinking Machines Lab released Inkling-Small, an open-weights model roughly a quarter the size of its flagship Inkling, and on the benchmarks that matter to developers it comes out ahead of its own teacher.
Jargon check. “Open-weights” means the actual trained model is downloadable and you can run it on your own hardware: no API, no subscription, no one watching. “Distilled” means a big model was used to train a small one to imitate it. Normally the student ends up a bit worse than the teacher. Not this time.
The numbers
Inkling-Small is a 276-billion-parameter Mixture-of-Experts model with just 12 billion parameters active at any moment. (Mixture-of-Experts, or MoE, means only a fraction of the model fires for each request, so it runs far cheaper than its total size suggests.) The full Inkling is 975B total, 41B active, so Small is about a quarter of the heft.
Yet it scores 31.6% on Humanity’s Last Exam, a brutal reasoning benchmark, ahead of Inkling’s 29.7%, and 80.2% on SWE-bench Verified (real GitHub bug-fixing tasks) against the bigger model’s 77.6%. It keeps the 1-million-token context window, native reasoning over audio and images, and adjustable “thinking effort”.
Why this matters more than another frontier flex
The weights are on Hugging Face under Apache 2.0, which is about as permissive as licences get: use it commercially, modify it, bundle it inside your product, no permission required. The full-precision checkpoint needs around 600GB of VRAM, but the compressed NVFP4 version drops that to 180GB, which is a serious workstation rather than a data centre.
This is the whole Thinking Machines thesis in one release. Murati’s bet, ever since she walked out of OpenAI, is that the future belongs to small, open models that companies fine-tune on their own data until they beat the frontier at their specific job, rather than one giant chatbot to rule them all. Inkling-Small is the cheap, downloadable proof of concept. You do not rent it. You own it.
The commercial hook is Tinker, Thinking Machines’ platform for fine-tuning these models on your own data. Give away the model, sell the tooling that makes it yours. It is the opposite of the Meta play landing in this same news cycle, and arguably the more honest one.
The catch
Thinking Machines is refreshingly upfront that these are not the strongest models on the planet. Claude Opus 5 and the top GPT-5.6 tier will out-reason them on general tasks. The pitch is control and cost, not raw horsepower. If you are a business that would rather run a capable model on your own tin than pipe your data through someone else’s API forever, that trade is starting to look very attractive.
Did you know: “distillation” borrows its name from chemistry, where you boil off a liquid to capture its purest essence. In AI the essence being captured is a giant model’s behaviour, poured into a container small enough to sit under your desk.
Related on Top Tool Stack: Meta’s Muse Code wants your code · Alibaba’s Qwen3.8-Max