Inkling 2 Small

Inkling 2 Small is Thinking Machines’ second-generation language model, built as a mixture of experts with 594 billion total parameters and 17 billion active parameters. It accepts images and supports variable reasoning effort. The company reports an Artificial Analysis Intelligence Index score of 31, alongside improvements in coding, vision, knowledge work, and reasoning. Its training combined more efficient pretraining with a large reinforcement-learning run, including hundreds of thousands of long-horizon coding tasks judged on both outcomes and problem-solving approach. The model is initially available with closed weights through Tinker; the company plans an Apache 2.0 release after external safety testing and red-teaming.

The demonstrations emphasize autonomous work over extended tasks: generating interactive 3D scenes and websites, and improving software through testing and verification. Benchmark results show strong but uneven performance: Inkling ranks near the top on several coding tests and solves an average of 54.25 of 69 GPU-kernel optimization tasks in a 10-hour evaluation, while competitors lead some coding and optimization measures. In an internal forecasting experiment, it independently assessed data, built an evaluator, and trained a Qwen model using supervised fine-tuning and other experiments. After 4.5 hours and $426 in Tinker costs, the resulting model more than doubled its forecasting score. The release also describes the model’s capabilities and limitations, but the available details do not provide a full account of its safety evaluation.