Thinking Machines unveils Inkling 2 Small: 594B params, weights still locked

The model is a Mixture-of-Experts design: 594B total parameters, but only 17B active at a time. It takes image input and offers a "variable effort" setting that lets users control how much reasoning it burns. Thinking Machines says a 292-million-parameter visual encoder built on the Perception Encoder handles the image side.

On the Artificial Analysis Intelligence Index it notched 31 points, a clear jump over its predecessor, with the company pointing to gains in agentic coding, image processing, knowledge tasks and reasoning. Its own "Coding and Autoresearch" table credits Inkling 2 Small with 25.4% on FrontierCode v2.0, 72.1% on DeepSWE v1.1, 86.3% on SWEBench Pro v2 (Hard), 81.6% on Vibe Code Bench v1.1 and 31.9% on PostTrainBench v1.2. The same table shows rivals ahead in places: DeepSeek-V4.1-Flash-1017 beat it on FrontierCode (28.5%), DeepSWE (77.0%) and Vibe Code Bench (87.7%), while Kimi K3 reached 88.2% on SWEBench Pro v2 (Hard).

Training was split into four pre-training stages. The first crammed 8.4 trillion tokens through NVIDIA GB300 NVL72 hardware in NVFP4 precision in 68 hours, followed by a reinforcement learning run spanning more than 17 million rollouts across 1,075 steps, 9 days and 2,048 GPUs.

For now the weights stay shut. Inkling 2 Small is served through Tinker first, with a Playground open from announcement day plus support for fine-tuning and sampling. Thinking Machines says it will publish the weights under Apache 2.0 once external safety testing and red teaming wrap up, and the Nov 5 post pointed to "the following week" — but as of Nov 9 no release has been announced. Checkpoints are set to be downloadable the same day the weights arrive.