On November 5, 2026, Thinking Machines Lab released Inkling 2 Small, its second-generation model. It is a mixture-of-experts (MoE) model with 594B total parameters and only 17B active per inference, supporting image inputs and variable effort. It scored 31 on the Artificial Analysis Intelligence Index, far ahead of the previous generation. More importantly, the lab has promised to release the full weights under the Apache 2.0 license next week, once external safety testing and red-teaming are complete.
The gains come from improvements across the whole training pipeline: more efficient pre-training plus one large, unified reinforcement-learning run. The model was trained with RL on hundreds of thousands of long-horizon coding tasks, spanning multiple agent harnesses and context-compaction strategies, with rewards judging how it approached a problem, not just whether unit tests passed. On vision, a 292M-parameter vision encoder handles image inputs, letting the model build websites and 3D designs from images and refine its output using visual feedback. In official demos, it wrote a PC emulator from scratch that boots a real Linux kernel, and built a city simulator in Three.js with power generation, budgets, and emergency management.
Benchmarks: top-two finishes, but not a clean sweep
In the official comparison, Inkling 2 Small posted 72.1% on DeepSWE, 86.3% on the hard split of SWE-Bench Pro, and 25.4% on FrontierCode, plus 81.6% on MMMU Pro and 83.8% on ScreenSpot Pro for visual reasoning — top-two finishes in several rows. But the same table shows DeepSeek-V4.1-Flash ahead on several coding benchmarks, and Kimi K3 beating it on individual items. One highlight is autoresearch: in 4.5 hours and $426 of Tinker spend, the model autonomously trained Qwen3.5-4B into a time-series forecasting model, more than doubling its forecasting score.
The rollout schedule is worth noting: the model first arrives on the lab's own Tinker platform with closed weights, offering trials, fine-tuning, and sampling, with full weights to follow next week after external safety checks. Once released, it is set to become the highest-scoring open-weights model on DesignArena. For developers, the real test comes after the weights land: the official benchmarks look strong, but stability on long-horizon tasks still needs to be proven by the community. Anyone curious can already try it on Tinker Playground.
