A 594B Model Ships Closed, Then Opens

Thinking Machines is testing Inkling 2 Small behind a wall first and releasing full weights a week later.

Thinking Machines Lab released Inkling 2 Small on 5 November 2026, a sparse mixture-of-experts model with 594 billion total parameters and 17 billion active per token, initially available on its Tinker platform with closed weights. Full weights are planned for release under Apache 2.0 the following week, after external safety and red-team testing. Separately, Bridgewater’s AIA Labs has reported two distinct Tinker projects: a financial information-screening model built on Qwen3-235B that scored 84.66% against GPT-5.5’s 78.2% at roughly one-fourteenth the inference cost, and ReViSQL, a text-to-SQL model built on Kimi-K2.6 that reached 92.97% under sixteen-candidate self-consistency rather than in a single pass.