4B model hits 61.5% on SWE-bench Verified without distillation

A 4-billion-parameter model has hit 61.5% on SWE-bench Verified without distilling from a larger frontier model. Only a year ago, this level of performance was limited to models ten times larger.

The model, called FrogNano, is built on Qwen3.5-4B and was trained with reinforcement learning on roughly 1,500 synthetic software-engineering tasks. But the data volume isn't the story. The story is how those tasks were selected.

**A curriculum that updates in real time**

Instead of feeding the model a fixed problem set, the system generates new tasks each training round: hard enough to teach something, not so hard they're unsolvable. As the model improves, the difficulty ceiling rises with it — what the researchers call "online task synthesis." The result is a curriculum that adapts to the agent's current ability, rather than forcing it to stall on trivial problems or give up on ones out of reach.

**A stripped-down tool interface added 29 percentage points to the base model**

The second breakthrough, no less significant, is a simplified tool interface. Moving to just five tools lifted the base model from 8.3% to 37.2% on the same benchmark, before any reinforcement learning. The implication is clear: much of the failure of small models on coding tasks comes from overcomplicating the action environment, not from a lack of knowledge. When the environment is clean, the small model knows far more than we assumed.

**Five RL rounds delivered the final score**

After the initial jump from the simplified interface, the researchers ran five reinforcement-learning rounds with the adaptive task synthesis. Each round added a few more percentage points, culminating in 61.5%. For context, far larger models only recently crossed the 60% threshold on this test after massive engineering efforts.

**Paper available on arXiv**

The work is published as a preprint at arxiv.org/abs/2609.07925 under the title "FrogNano: Training a 4B Coding Agent via Online Task Synthesis." As a preprint, it has not yet undergone formal peer review, but the methodology is detailed enough for other groups to replicate and verify the claims.