Thinking Machines released an open-weights model, and an any-to-any MoE under Apache 2.0
Inkling is the lab's open-weights release. Separately, on 31 July, a mixture-of-experts system taking image and audio inputs shipped under Apache 2.0 — a permissive licence, in a month when the direction of travel has been toward bespoke terms.
Apache 2.0 is the notable part. In the last three weeks the sector produced a revenue-share licence on a 2.4T model, a bespoke instrument named after the model it governs, and now a plain permissive grant on a multimodal mixture-of-experts. The spread is the story.
An any-to-any architecture accepting image and audio inputs through a mixture-of-experts is a genuinely useful shape for downstream work: it is the kind of component teams build products on rather than benchmark against, which is exactly the case where licence friction decides adoption.
Thinking Machines' position gives the choice weight. A lab founded on a public argument about how AI research should be conducted, releasing under the most permissive standard terms available, is making that argument in the only form that is hard to discount.
Whether it holds as the lab's models get more valuable is the open question — and it is the question every lab that started permissive has eventually faced. Permissive terms have generally been a strategy for entering a market rather than for holding a position in one.
Thinking Machines Lab — Inkling: Our Open-Weights Model → · The Open Weights — Open-source AI, tracked daily →