Aztec-Coder-4B is a 4-billion-parameter agentic coding model from San Diego State University's James Silberrad Brown Center for AI Research (JSBCAI), fine-tuned from Qwen3.5-4B to investigate bugs, edit files, run commands, and verify its own fixes in real software repositories — a capability class the authors note has typically required 27B+ models — while fitting on consumer hardware (~8GB VRAM in BF16, ~5GB as an NVFP4 quantized variant, which retains roughly half the generalization capability at 38.0%/12-32 on the same evaluation slice). It was trained in three stages: seed demonstrations from ~1,875 test-verified coding trajectories generated by GLM-5.3, reinforcement learning via 145 batches of on-policy GRPO over 237 curated software-engineering problems using test-pass/fail as the only reward signal, and repeated generalization checks against unseen bugs, alongside general instruction-following data drawn from NVIDIA's Nemotron-Post-Training-Dataset-v2. On 121 held-out, never-trained-on real bugs verified by running each project's hidden test suite, it jumped from 10.1% (base Qwen3.5-4B) to 82.9% solve rate, while also improving general capabilities rather than trading them off — IFEval rose from 84.66 to 87.21 and MMLU-Pro from 64.0% to 70.0% — alongside gains on Live-60 real-world engineering tasks (15.0% to 21.7%), with Terminal-Bench 1.0 held flat at 33.8% and Terminal-Bench 2.1 results still pending. It's served via vLLM with the Qwen3.5 chat template, qwen3 reasoning parser, and qwen3_coder tool-call format, is tuned specifically for sandboxed container agent loops rather than general deployment, inherits its safety behavior unmodified from the base model (the RL phase optimized purely for test-passing with no safety-specific training), and is released under Apache-2.0.
