Benchmarks & research
Mica-v0.1-4B is an open-weight decision model based on Qwen3.5-4B that speaks the TypeSafe API format.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
Mica handles up to 8,192 input tokens, or 32,768 in v0.5, and returns probabilities without generating text. The maker claims strong robustness against planted instructions. That claim sits alongside weaker results than Jev on knowledge tasks and long English policies.
Mica’s local server can receive the decision requests your software would otherwise send to TypeSafe. Give it a situation, questions and the allowed answers, and it estimates each answer’s likelihood. Those are Mica’s answers rather than Jev’s. Reported performance varies by language, task and model version.