We build smaller models—and make them prove they deserve to exist.

Anthropic 2 studies a practical question: when should capability that lives in traces and large models become a smaller portable model?

Our first answer is TinyRamp, a 0.5B finance generalist. Our second is Distillery, the system that curates the data, selects a method, trains the candidate, and decides whether the result is actually worth deploying.

We publish the model claim, the system, and the evaluation separately so each can be challenged on its own terms.