Mira Murati's lab shipped a giant open model — and told you to make it smaller

The default enterprise reflex is to reach for the biggest frontier API and prompt it into shape. Inkling is a wager against that reflex: that for the specific work a team does every day, a smaller model you fine-tuned on your own data and own outright will beat a general-purpose giant you rent. Whether or not Inkling is the model that proves it, that bet is the thing worth reading.

Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, released Inkling on 15 July — its first model, and a notable one. It is a mixture-of-experts transformer with 975B total parameters and 41B active per token, a context window of 1,048,576 tokens, and pretraining across 45 trillion tokens of text, images, audio, and video. It takes multimodal input and produces text. Crucially, it ships under Apache 2.0 — a fully permissive licence that allows commercial use, modification, and redistribution with no field-of-use strings — which makes it the largest open-weight model to come out of a US lab.

The release is deliberately paired with two things that tell you what it is for. First, Inkling-Small: a preview of a 12B-active variant trained on a similar recipe that, per the lab, lands close to the larger model's quality at roughly a quarter of the size, with the lower cost and latency that implies. Second, Tinker — Thinking Machines' customisation platform — is the front door: Inkling is available to fine-tune there on day one, alongside a straight weights download from Hugging Face. The framing is not 'here is a model to call.' It is 'here is a base to adapt.'

The benchmarks make the point better than a press line could. Inkling posts a strong 79.8 on IFBench for instruction-following against GLM 5.2's 73.3, but loses to Kimi K2.6 and GLM 5.2 on Terminal Bench 2.1. This is not a model launched to top a leaderboard, and it is not pretending to be. Thinking Machines' own thesis is that AI which organisations can adapt for themselves will outperform the one-size-fits-all models the biggest labs currently sell — because most real-world problems are not solved well even by the best generalist, and the gap is closed by fine-tuning on knowledge only the customer has.

That reframes what 'good' means. A team fine-tuning for one narrow task cares less about a model's max-effort score on a public eval and more about efficiency: accuracy on their workflow, at their latency, at their cost. An open base under Apache 2.0 plus a customisation platform is a coherent product for that buyer in a way a metered frontier endpoint is not — you can specialise it, run it where you like, and you are not exposed to a vendor changing the model, the price, or the policy underneath you.

For three years the industry's centre of gravity has been the generalist frontier model, and the assumed path to production was better prompting and retrieval on top of someone else's API. Inkling is one of the clearest signals yet that a second path is hardening: take a capable open base, fine-tune it into a specialist, own the weights, and deploy it on your own terms. It is the same instinct visible across the open-weight wave this summer — the difference here is a well-funded lab building its entire go-to-market around adaptation rather than scale.

The small-model half of the release is the quiet tell. If a 12B-active variant can approach the big model on the task you actually care about once it is tuned, the economics of 'own and specialise' beat 'rent and prompt' for a growing share of production work — especially anything high-volume, latency-sensitive, or bound by where data is allowed to live.

The move is not to fine-tune everything — most tasks are still cheaper solved by prompting a hosted model, and fine-tuning only earns its keep when a task is narrow, repeated, and stable enough that a tuned small model pays back the effort. The move is to know which of your workloads are that shape, and to test one honestly: take a task where you have proprietary data and a clear success metric, tune a small open model on it, and measure quality, latency, and cost against your current API call. If the specialist wins on your metric, you have a cheaper, portable, self-owned system; if it does not, you learned that for the price of an experiment.

The discipline that decides the outcome is the boring part: a real evaluation set drawn from your own work, a holdout you did not train on, and a clear-eyed read on whether the tuned model actually beats the generalist or merely matches it more cheaply. An open base makes the experiment possible; only your own evals tell you whether specialisation paid off.

Inkling's headline is a giant open model from a famous founder. Its argument is the opposite of giant: that the frontier is table stakes and the edge now lives in models teams shape for themselves, own outright, and run where their data has to stay. Not every workload wants a fine-tuned specialist — but the ones that are narrow, repetitive, and yours are exactly the ones that have been quietly overpaying to rent a generalist.