Inkling 2 Small reaches Tinker with its weight release still conditional
Access for inference and fine-tuning precedes the planned weight release. A forecasting example shows why evaluation design deserves separate scrutiny.
An agent improving its own evaluation score raises a separate question: does the improvement survive an independent test? Thinking Machines Lab's Inkling 2 Small announcement, dated November 5, 2026, includes a forecasting experiment that makes this distinction unusually visible.
The agent discovered that its evaluator reused a random seed, producing identical forecast samples. Later, a reinforcement-learning attempt made the forecasts overconfident and reduced their score, so the agent stopped that run. The final checkpoint improved on unseen data, but by less than its development evaluations suggested.
Availability also needs careful reading. The announcement offered inference and fine-tuning through Tinker while keeping the weights closed. An Apache 2.0 weight release was planned after external safety testing; that statement does not establish that distribution has happened.
For an application evaluation, I would assess the scoring procedure separately from the model being scored. The published example documents useful corrective behavior within that experiment. It leaves open whether the same ability to diagnose evaluation errors transfers to other research settings.
Source: Thinking Machines Lab, Introducing Inkling 2 Small.