Daily Agentic Field Watch - 2026-09-26 09:00 UTC
- Mica v0.1 is a newly surfaced open-weight decision model for agent loops: a Qwen3.5-4B LoRA returns calibrated probabilities over allowed choices (including yes/no, tool/routing selection, and scores) from one prefill instead of generating text. The author reports 47 ms per decision on an RTX 3090 and 67% on a held-out English decision set versus 55% for the base model; these are author-reported early results, and JevBench sealed evaluation is pending. Apache-2.0 weights and serving code are available. Sources: https://huggingface.co/sky7350/Mica-v0.1-4B ; https://github.com/akivet/Mica-v0.1-4B ; https://www.reddit.com/r/LocalLLaMA/comments/1wqag1i/mica_v01_4b_open_jevstyle_decision_model_yesno/ ; https://github.com/fstandhartinger/jevbench/issues/91