Mira v12 (System One package 0.1.5)
Small open typed-decision model (140M): choice / score / noul with calibrated
probabilities, CPU-first, Apache-2.0. Hugging Face: sagea-ai/Mira-v1.
Measured (2026-10-02, box-01 CPU, native code)
- Neutral (typed-decisions rev c76749ec, 2000 Q): acc 0.6125 [0.589, 0.634], ECE 0.083, Brier 0.495 — choice 0.688 / score 0.618 / noul 0.530.
- Pilots: AGNews 0.23, Emotion 0.87, Banking77 0.93.
- Latency: same architecture as 0.1.4 (per-option cross-encoder).
Limits (honest)
Neutral accuracy trails Julia-1 (0.726) and Laya (0.759). Strengths are calibration (ECE 0.083 vs 0.2+ elsewhere), CPU speed, size, and support-style routing (grid 0.745 lineage). Conventions (v11+): choice encodes "name: description"; score levels read anchored ("level i of n-1: ...").
Run: from mira_infer import MiraDecider; MiraDecider(...).decide(state, questions).