Thinking Machines Lab debuts 594B Inkling 2 Small MoE on Tinker with Apache 2.0 weights planned

Thinking Machines Lab introduced Inkling 2 Small on 5 November 2026 as the first release in its second-generation model family. Built on a Mixture-of-Experts (MoE) architecture with 594 billion total parameters and 17 billion active parameters per token, the multimodal model accepts image inputs and supports configurable reasoning effort. Inkling 2 Small is live for interactive use on Tinker Playground and for sampling and fine-tuning on the Tinker platform, with full open weights and downloadable checkpoints scheduled for release under an Apache 2.0 licence next week after external red-teaming concludes.

Previously tested in the wild on OpenCode under the stealth name Fledge-Alpha, Inkling 2 Small was trained with a large unified reinforcement learning run across hundreds of thousands of long-horizon coding and autoresearch tasks, using an agentic RL grader that scores problem-solving approach alongside unit-test pass rates. Incorporating visual feedback loops, the model built a procedural Three.js city simulator without pre-made assets, a browser-booted Python Linux PC emulator and iterative Blender 3D scenes.

Key numbers

• 31 on the Artificial Analysis Intelligence Index, alongside 86.3% on SWEBench Pro v2 (Hard) and 72.1% on DeepSWE v1.1
• 81.6% on Vibe Code Bench v1.1, 31.9% on PostTrainBench v1.2 and 25.4% on FrontierCode v2.0
• 54.25 out of 69 mean GPU kernel optimization tasks solved (494.27 task-hours AUC; 64/69 Pass@4) over a 10-hour PMPP-Hard budget
• 6.02 times speedup on DeepSeek-V3.2 sparse attention forward-and-backward kernel autoresearch over a 10-hour session