On November 5, Thinking Machines Lab introduced Inkling 2 Small, its second model generation. It is a Mixture-of-Experts model with 594 billion total parameters and 17 billion active, it takes image inputs, and the effort it spends is adjustable.
On the Artificial Analysis Intelligence Index it scores 31, well above anything the lab had before. Artificial Analysis puts it level with Gemini 3.8 Flash (high) and writes that, once the weights are out, it would be the strongest open model outside China.
Open, but not yet
For now Inkling 2 Small runs through Tinker with closed weights. The lab intends to publish the full weights next week under Apache 2.0, after external safety testing and red-teaming are done. You can try the model in Tinker Playground today, fine-tuning and sampling run through Tinker, and the checkpoints become downloadable when the weights land.
That order was announced in advance, and this is the first time the lab has walked a frontier model through it: API first, then the testing, then the weights. Wait for the license and you lose a week. Don’t wait, and you are working with a closed model while calling it open.
How it was trained
The lab credits the gains to improvements at every stage, including more efficient pretraining and one large unified RL run. The interesting detail sits in the RL: training covered hundreds of thousands of long-horizon coding tasks, across multiple agent harnesses and multiple context-compaction strategies.
Thinking Machines also stopped hanging the reward on unit tests alone. Many tasks carried an agentic grader that scored how the model approached the problem, not just whether the code passed. Early checkpoints ran under the name Fledge-Alpha to measure coding performance on real work.
The lab describes the result as better behavior inside agent loops rather than a better benchmark number: the model follows instructions, considers changes carefully, and verifies its own work unasked. Add vision and agentic capability, which let it fold visual feedback into building websites, games and 3D designs; as evidence the lab shows a Three.js city simulator with procedurally generated buildings and a from-scratch PC emulator in Python that boots a real Linux kernel.
The competition Claude Code can actually feel
If you build agents, this is the point where open weights aim at the same class as the closed ones for the first time: 17 billion active parameters are cheap to serve, and the rest is routing. If the Apache license really arrives next week, what moves is not the top score but the price of getting close to it.