Mica v0.1 4B Released: Open-Source Decision Model for Agent Loops Trained for Under $30

Mica v0.1 4B is a new open-source decision model designed for AI agent loops, such as gating, routing, and determining whether to ask the user or act. Released under the Apache-2.0 license, it outputs calibrated probabilities for yes/no answers, multiple-choice selections, or scored levels without generating text. The model is compatible with the TypeSafe /v1/systemone format, allowing integration with existing Jev-style workflows.

Architecture and Training

The model is built on Qwen3.5-4B with a rank-16 LoRA applied to all 32 layers, merged into a standard checkpoint. Training utilized approximately 34k source decisions expanded to 77,732 rows (34.7M tokens), split evenly between English and Korean across domains like coding agents, code review, and policy rules. The entire training process, including experiments, cost less than $30 in rented GPU time on RTX 3090s.

Performance Benchmarks

On a held-out set of 7,328 decisions, Mica scored 67.0 on the English subset, outperforming JevK5 4B (61.0) and Kev 4B (57.0), though trailing the closed API Jev 1.13 (74.1). In speed tests on an RTX 3090, the Q4_K_M quantized version achieved a median latency of 47 ms, significantly faster than Kev 4B (76 ms) and JevK5 4B (99 ms).

Key Features and Limitations

• Robustness: Mica maintains 69% accuracy against in-data prompt injection attacks, compared to 18% for Jev.
• Calibration: High-confidence predictions (≥0.9) have a low error rate of 2.5% (ECE 5.4%).
• Efficiency: The Q5_K_M GGUF file is 3.5 GB with negligible accuracy loss compared to BF16.
• Limitations: It struggles with knowledge-heavy tasks (MMLU-Pro score of 53.0 vs. 82.3 for Jev) and long English policy documents. It also currently lacks strong differentiation between reversible and irreversible actions.