Microsoft releases FrogNano-4B, an open-weight 4B coding agent trained only with RL
Built on Qwen3.5-4B under Apache 2.0 and trained with reinforcement learning on about 1,500 synthetic repo tasks (no stronger-model trajectories), FrogNano scores 61.5% on SWE-bench Verified and 37.6% on SWE-bench Pro in Microsoft's own five-tool Leaf harness at ~131K context. Microsoft calls it a research artifact: Python-heavy, BF16 weights need about 9.3GB, and quantized builds aren't validated yet.