Thinking Machines Lab

US AI lab, and as of 2026-07-15 an open-weight one: it shipped Inkling, a 975B/41B Apache-2.0 multimodal MoE, plus a 276B/12B Inkling-Small in testing inkling-announcement.

Position on the map

A US open-weight lab that is neither a hyperscaler nor a Chinese lab — a slot the provider map didn't have. The open field here has been qwen, deepseek, z-ai and mistral-ai, with google hedging via dual-track gemma-4. Willison frames Inkling as joining NVIDIA Nemotron and Gemma 4 in a thin US open-weights ecosystem inkling-willison-review.

The lab doesn't claim the frontier — it says so in the announcement — and instead sells Tinker, a fine-tuning platform. The weights are free and the adaptation is the product. Compare mistral-ai, which also moved up-stack from open weights to a platform, but kept a proprietary frontier tier alongside; Thinking Machines skips the frontier claim entirely.

Caveat

Written 2026-07-16 as: "One release, one page. Everything here rests on release-day material — the lab's own announcement (T1) and one practitioner review (T2). No independent benchmark re-run, no track record, no pricing sheet."

Re-checked 2026-08-30, and two thirds of that is no longer true.

• There is a pricing sheet. Tinker bills per million tokens across three meters — prefill, sample and train — rather than per GPU-hour. Fine-tuning Qwen3-8B was $0.40 per million training tokens, raised to $0.44 on 2026-07-17; evaluation sampling adds $0.40–$0.60. That is a rate card, and it puts the lab on the same axis as every other entry in llm-api-pricing — except that what it sells by the token is adaptation, not inference.
• It is no longer one release. Inkling-Small (276B/12B) shipped out of testing, and the lab has since announced Interactivity Research Grants — multiple $100,000 awards plus $25,000 in Tinker credits. The stated reason: the lab says the open research community is the party positioned to produce independent, fair evaluations. A lab paying for its own third-party evaluation is a different governance shape from one waiting to be evaluated.
• Still no independent benchmark re-run and no track record. The Tinker cookbook ships a 12-benchmark framework with scores verified against published results, but that is the lab's own harness on other people's benchmarks. Nobody outside has re-run the models. Corrected 2026-10-06: the Inkling announcement itself says six of its scores (Humanity's Last Exam, GPQA Diamond, GDPVal, Tau 3 Banking, AA Omniscience, MMMU Pro) are "the score reported by Artificial Analysis", so a third party did run the model on those evals inkling-announcement. What is still missing is an outside re-run of the lab's own benchmarks, such as the TimeSpeak figures for the interaction model.

Related

inkling · tinker · inkling-announcement · inkling-willison-review · open-weight-models · llm-provider · mistral-ai · google