Microsoft's FrogNano: 4B Coding Agent Trained with Pure RL Hits 61.5% on SWE-bench

Microsoft Research Montreal's Froggy team, together with Mila and UC San Diego, released FrogNano, a coding agent with only 4B parameters. Built on Qwen3.5-4B, its post-training relies entirely on pure reinforcement learning over software engineering tasks synthesized online by TaskPilot—no distillation, no teacher-trajectory SFT, and no human-labeled data—yet it reaches 61.5% on SWE-bench Verified, challenging the common assumption that small models must be distilled from frontier models.

Confirmed

Model and base: FrogNano is based on Qwen3.5-4B, positioned as a coding agent that runs on low-spec machines
Training: pure RL post-training with zero distillation and zero teacher-trajectory SFT; tasks are synthesized online by TaskPilot, roughly 1,500 environments/tasks total (5 iterations × 300 tasks, then updates stopped)
Result: 61.5% on SWE-bench Verified
Design constraints: both a frontier base model and a frontier teacher model are excluded, aiming to lower the cost and deployment barrier of frontier SWE agents; disclosed technical details indicate the adaptive synthesis of RL environments is itself a core part of training

Why it matters

It challenges the common practice that small models must be distilled from frontier models, offering new evidence that competitive coding agents can be trained cheaply
dairai called it "a big deal" that a competitive small coding agent can be trained without frontier-model distillation, showing strong community interest

2026-09-09 ~ 2026-09-11 · 7 related posts

Primary sources

Microsoft's FrogNano trains a 4B coding agent almost entirely with RL, no teacher model — ossm-me · 2026-09-10 21:46 (UTC+08:00)
FrogNano details: adaptive on-policy synthetic RL envs train a 4B coding model to 61.5% — sivareddyg · 2026-09-09 13:04 (UTC+08:00)
Zero-Distillation 4B Model Hits 61.5% on SWE-bench via Pure RL — rajammanabrolu · 2026-09-11 04:24 (UTC+08:00)

FrogNano: 4B model hits 61.5% on SWE-bench Verified with pure RL, zero distillation — sivareddyg · 2026-09-09
[source] FrogNano details: adaptive on-policy synthetic RL envs train a 4B coding model to 61.5% — sivareddyg · 2026-09-09
Microsoft's FrogNano: 4B coding agent trained with pure RL, no frontier distillation — dair_ai · 2026-09-09
FrogNano: a 4B model trained purely with RL on synthetic tasks hits repo-level coding — burkov · 2026-09-10
[source] Microsoft's FrogNano trains a 4B coding agent almost entirely with RL, no teacher model — ossm-me · 2026-09-10
[source] Zero-Distillation 4B Model Hits 61.5% on SWE-bench via Pure RL — rajammanabrolu · 2026-09-11

1 near-duplicate retellings: DimitrisPapail