Thinking Machines Lab debuts Inkling 2 Small, a MoE model with 594B total and 17B active parameters, and plans to release weights next week after safety testing (Thinking Machines Lab)

Thinking Machines Lab has unveiled Inkling 2 Small, a new Mixture-of-Experts model boasting 594B total parameters and 17B active parameters. This latest model boasts impressive capabilities, excelling in long-horizon autonomous work, image input processing, and variable effort. Inkling 2 Small has outperformed previous models on the Artificial Analysis Intelligence Index, scoring 31, reflecting improvements in agentic coding, vision, knowledge work, and reasoning.

The model's enhanced features stem from improvements throughout its training process, including more efficient pre-training and a large, unified reinforcement learning run. Inkling 2 Small will be available initially through Tinker, with closed weights following the lab's safe path to open weights approach. The full weights will be released under an Apache 2.0 license next week after external safety testing and red-teaming.

Inkling 2 Small demonstrates strong visual capabilities, using a 292M-parameter vision encoder to process image inputs. The model can interpret and validate visual coding output, generate high-quality deliverables, and even outperform human researchers in certain tasks such as wine tasting. It improves upon its predecessor by following instructions better, considering changes more carefully, and verifying its work.

Moreover, Inkling 2 Small has been trained on challenging STEM problems, instruction following, and forecasting tasks. It adapts quickly to post-training scenarios, achieving the final performance of Inkling Small after just 75 training steps. The model is poised to become the top open-weights model on DesignArena and has shown strong performance in finance and quantitative reasoning tasks.