Thinking Machines Lab

The debut model from the lab founded by OpenAI's former chief technology officer is built on a Chinese architecture, trained partly on another Chinese lab's output, and given away under Apache. thinkingmachines/inkling, listed 17 July 2026, is 975B total, 41B active — a 66-layer decoder-only transformer routing each token to 6 of 256 experts, plus 2 shared experts active on every tokensource, accessed 2026-08-28, published under Apache License 2.0source, accessed 2026-08-28. Wikipedia's account of the release records that it incorporated architecture from the Chinese model DeepSeek-V3 and synthetic data from Moonshot AI's Kimi K2.5source, accessed 2026-08-28. So a DeepSeek architecture and Moonshot AI training data sit inside the first model shipped by an American company that raised US$2 billion at a US$12 billion valuation (July 2025), led by Andreessen Horowitz with Nvidia, AMD, Cisco and Jane Street participatingsource, accessed 2026-08-28.

Everything the lab has listed arrived inside thirteen days. Five rows — thinkingmachines/inkling, its batch and free variants, thinkingmachines/inkling-small and that model's free variant — were listed between 17 and 30 July 2026, and nothing has followed; one of them, the batch row, has since left the router. Inkling Small is a 276-billion-parameter distillation with 12 billion active. Every model this lab has published is downloadable, and every one is also available at no charge through the router: two models, four rows still listed, no closed weights and no paywall that cannot be walked around.

For two months there were three ways to call the same model, and they did not rank the way the names suggest. thinkingmachines/inkling:free serves the full window at no cost, and the standard row serves the same 1048576tokensopenrouter-models, last checked 2026-09-24. The batch row served half of that window, and in the router's catalogue as captured on 28 August 2026 its input rate headed five cents per million tokens above the standard row's — a batch listing above the row it batches, where the convention is a discount. Neither rate was necessarily this lab's: each was the top listed provider's rate for its row, so the inversion was a fact about who headed each separately ranked row, not a decision the lab made, and by 30 August the standard row's input rate had risen to match. The batch route is now history. thinkingmachines/inkling:batch was last seen in OpenRouter's model catalogue on 21 September 2026 and was gone by the 23rd; on 25 September the catalogue listed the standard and free Inkling rows but no batch one, and the batch ID's endpoint lookup returned no providers.