Mica turns a 4B model into a fast action judge, not a chatbot
Mica v0.1 lets an agent choose actions without generating text, making a 4B model usable as a low-latency decision layer. Give it a state, a question, and allowed answers, and it returns probabilities or scores in one prefill, with support for 2 to 255 choices. On JevBench's 231 public items, it scored 100% on easy tasks, 100% on original tasks, and 64.9% on hard tasks through its llama.cpp server.