Thinking Machines Lab released Inkling 2 Small on Nov 5. You can try it in the Tinker Playground today, and sample or fine-tune it on Tinker. Checkpoints become downloadable when the weights are released, which the company says it intends to do next week once red-teaming is done.
The company reports a score of 31 on the Artificial Analysis Intelligence Index. On coding, its own table shows 25.4% on FrontierCode v2.0, 72.1% on DeepSWE v1.1 and 86.3% on SWE-bench Pro v2 (Hard). It is second-highest in those rows, behind DeepSeek V4.1 Flash on the first two and Kimi K3 on the third. On vision and several finance benchmarks it leads the open-weight models it compared against. These are the lab's own numbers.
In one demo, the model ran an autoresearch task on its own: it trained Qwen3.5-4B on Tinker to forecast time series, spent about 4.5 hours and $426 of a $500 budget, and more than doubled the forecasting score. The post also admits limits: it isn't the most succinct writer, doesn't always pick the best tools in repos with many skills, and the chat experience is less polished.