Microsoft Research released FrogNano, reporting a 4-billion-parameter software-engineering agent trained entirely through reinforcement learning on synthetic tasks, without distillation from
Microsoft Research introduced FrogNano, a 4-billion-parameter coding agent trained purely with reinforcement learning on synthetic software-engineering tasks—without distillation from larger frontier models. Its TaskPilot pipeline generates tasks near the model’s “frontier of learnability” across roughly 1,500 environments. After five 300-task iterations, FrogNano reportedly reached 61.5% on SWE-bench Verified, suggesting smaller, cheaper agents can rival systems many times their size.