Microsoft open-sources FrogNano-4B, a repo-level coding agent model trained with RL on 1,500 synthetic SWE tasks
jacek2023 · reddit · 2026-10-03
Microsoft released FrogNano-4B-2609 on Hugging Face, a compact agentic coding model aimed at "GPU-poor" users.
- Derived from Qwen/Qwen3.5-4B, inheriting its 32-layer hybrid Gated DeltaNet + gated-attention architecture, with text-only post-training focused on repository-level software engineering
- RL training runs on 1,500 synthetic SWE task environments generated and calibrated via TaskPilot, using the five-tool Leaf harness with executable-test rewards over full multi-turn coding trajectories
- No behavioral distillation: it doesn't learn from stronger models' trajectories, reasoning, or patch targets; goals are long-horizon repo navigation, debugging, code editing, and patch generation
- Caveats: sensitive to Leaf harness and test quality, Python/English-heavy training data, and generated patches may be incorrect or insecure, requiring human review and regression testing
More from coding & agent
- PhD-turned-founder: LLMs flipped which criteria kill programming tool startups — jimmykoppel · 2026-10-03
- Veteran engineer: AI coding now beats humans on quality, not just speed — facontidavide · 2026-10-03
- Television: an open-source GUI that gives your AI agent a visual workspace for artifacts — genmon · 2026-10-03
- Swarms: a 7.2k-star enterprise multi-agent orchestration framework on GitHub — KyeGomezB · 2026-10-03
- CMU blog: harness engineering boosts scientific agents without retraining models — niloofar_mire · 2026-10-03
- Swarms v16 'OverClock' ships 114 commits with MCPDeployer and DecisionModel — KyeGomezB · 2026-10-03