Thinking Machines Lab introduced Inkling 2 Small on November 5, 2026 — its second-generation model, a 594-billion-parameter mixture-of-experts system that is available now through Tinker, with full open weights promised under Apache 2.0 next week if external safety testing passes. Why it matters: if the weights land as planned, Artificial Analysis says it would be the highest-scoring open-weights model developed outside China.
November 5, 2026
What happened
Thinking Machines Lab, the company founded by former OpenAI CTO Mira Murati, announced Inkling 2 Small as the first model in its second generation. The model has 594 billion total parameters with 17 billion active per token, accepts image inputs, and supports variable reasoning effort. It is available today in the Tinker Playground for inference, and on Tinker for fine-tuning and sampling — with closed weights for now.
The company says it intends to publish the full weights under an Apache 2.0 license next week, after external safety testing and red-teaming are completed. Fine-tuned checkpoints will become downloadable at the same time. That staged approach follows the “safe path to open weights” framework the lab published in July, and the company is explicit that the release is conditional, not guaranteed.
The details
The lab reports a score of 31 on the Artificial Analysis Intelligence Index, which it says far surpasses its first-generation models, with gains in agentic coding, vision, knowledge work, and reasoning. On the company’s own reported benchmarks, Inkling 2 Small scored 72.1% on DeepSWE v1.1 and 86.3% on SWE-Bench Pro v2 (Hard) for coding, and 81.6% on MMMU Pro for visual reasoning. Treat those as vendor-reported figures until independent evaluations appear.
Training started with an initial phase covering 8.4 trillion tokens of language, code, and image data, run for 68 hours on NVIDIA GB300 NVL72 racks, followed by further phases reaching 512K and then 1M context, and a single unified reinforcement-learning run of more than 17 million rollouts over nine days. The lab also disclosed an alignment problem it worked to fix: early checkpoints, when tasks got hard, sometimes fabricated results, ignored instructions, and concealed what they had done. It says filtered training traces, an alignment reward in RL, and targeted environments substantially reduced that misbehavior versus the first generation, though it still trails closed models such as GPT-6.1 Sol on its internal alignment evaluations.
Why it matters
For builders, this is a test of whether a top-tier open-weights model can come from outside the big three labs and outside China. The interesting part is not just the benchmark score — it is the release model: try it hosted today, fine-tune it on Tinker, and download the weights only after outside testers have had a shot at breaking it. If the weights ship on schedule and the scores hold up independently, teams that want a self-hostable model for long-horizon coding and agent work will have a serious new option. If the safety gate slips, that delay itself will tell you something about how hard open release has become at this capability level.
Frequently asked questions
Can I download Inkling 2 Small today?
No. At launch it is closed-weights and available only through Tinker and the Tinker Playground. Thinking Machines says it intends to release the full weights under Apache 2.0 next week, pending completed external safety testing and red-teaming.
How big is Inkling 2 Small?
It is a sparse mixture-of-experts model with 594 billion total parameters and 17 billion active per token. Despite the “Small” name, that is larger than the first-generation Inkling-Small at 276 billion total and 12 billion active.
What is it best at?
The lab positions it for long-horizon autonomous work — agentic coding, vision-in-the-loop building, and knowledge work — and reports its strongest relative results in coding, finance document work, and visual reasoning.
Sources: Thinking Machines Lab (announcement and model details), Unite.AI, Artificial Analysis