FrogNano is a compact repository-level coding agent from Microsoft, derived from Qwen3.5-4B and post-trained exclusively with reinforcement learning — no stronger-model trajectories or distillation targets — on roughly 1,500 synthetic software-engineering tasks generated, validated, and calibrated against the evolving policy using TaskPilot. It operates through the lightweight five-tool "Leaf" harness, iteratively navigating repositories, editing code, running shell commands and tests, and producing candidate multi-file patches from natural-language issue descriptions, supporting a combined interaction context of roughly 131K tokens and up to 8,192 generated tokens per turn. Across five TaskPilot-guided RL iterations, its SWE-bench Verified Avg@3 resolution rate climbed from the base Qwen3.5-4B's 39.4% to 61.5% — a 22.1-point, ~56% relative improvement — with held-out results of 37.6% on SWE-bench Pro, 31.1% on Terminal-Bench 2.0, and 47.3% on PatchEval-Verified (plus Pass@3 scores of 71.0% and 47.6% on SWE-bench Verified and Pro respectively). The model is explicitly English-language and Python-heavy in its validated scope, inherits but does not support image/video input despite architectural remnants from the base model, is not independently safety-aligned beyond its functional-correctness RL training, and requires sandboxed execution plus qualified human review, regression testing, and security validation of any generated patch before use; it's released under the MIT license (noted separately as Apache 2.0 in the model table), runs on SGLang/vLLM/Transformers-compatible infrastructure, and is positioned as a research artifact for studying repository-level coding agents rather than a production-ready system.
