Thinking Machines launches Inkling 2 Small, a 594 billion-parameter open-weight mixture-of-experts model
Thinking Machines, the startup founded by former OpenAI CTO Mira Murati, has released Inkling 2 Small, its second-generation open-weight mixture-of-experts model. It has 594 billion parameters in total, with 17 billion active at once, takes text and images, and is framed as a fine-tuning base with adjustable reasoning depth so users can trade speed for thoroughness. The company calls it the strongest Western open-weight model.
On the Artificial Analysis Intelligence Index it scores 31, 18 points above Inkling Small’s 13 and roughly level with Google’s Gemini 3.8 Flash. Agentic scores rise sharply: AutomationBench-AA from 5% to 46%, FrontierCode 2.0 from 5% to 25% (25.4% in the company’s figures), and AA-Briefcase to 1,527 Elo, 615 above the predecessor. Hallucination falls from 61% to 27%, while the share of questions it answers drops from 77% to 53%. Average output is about 150,000 tokens per task, roughly four times the predecessor, at $1.18 per task—nearly six times as costly, but still well below Gemini 3.8 Flash ($7.77) and Claude Sonnet 5.5 ($10.02). Training used about 12 trillion tokens overall; the post-base reinforcement-learning stage used 3 trillion tokens on 2,048 GPUs over nine days, including hundreds of thousands of long-horizon coding tasks.
Further reported scores include DeepSWE 72.1% and SWEBench Pro v2 (Hard) 86.3%; Deepseek-V4.1-Flash still leads several of those tests (28.5% FrontierCode, 77% DeepSWE). Vision strengths include GDP.pdf 25.2%, MMMU Pro 81.6%, PerceptionBench 64.7%, and 1,340 Elo on DesignArena Website, plus a wine-tasting demo of 19 of 20 correct IDs from photos. Limits include inefficient communication, weaker tool choice on large codebases, no audio, weaker chat fit, overreaction on health topics, and cluttered output. Weights are slated for Apache 2.0 after safety testing; access is via Tinker, with more family sizes planned.
Continue in the app — vote & join in ➔