AIFrogNano coding agent reportedly reaches 61.5% on SWE-bench Verified
A post summarizing the FrogNano paper describes a 4-billion-parameter agent trained on about 1,500 synthetic software-engineering tasks that adapt as the model improves, without frontier-model distillation.
TLDR
A post summarizing the FrogNano paper says the agent starts from Qwen3.5-4B and uses reinforcement learning with synthetic tasks chosen to remain challenging but learnable as it improves. The post reports that switching to a simpler five-tool interface raised the base model’s SWE-bench Verified score from 8.3% to 37.2%. After five rounds, FrogNano reportedly reached 61.5%, without frontier-model distillation.
