9B model jumps to 56.4% on SWE-bench with 6K examples

A 9B model jumped from 41.8% to 56.4% on SWE-bench Verified after reinforcement learning on 6,000 examples, per the Agent Lightning paper on arXiv. That's a 14.6 point gain, not from a frontier lab's flagship.

The number that matters isn't 56.4%. It's 6,000. That's a dataset size a small team can build without a research budget. The paper ships with the full training and workflow scripts, not just results. Qwen3 is the base model, and the whole setup runs on roughly 3,500 lines of training code.

This is the difference between agentic RL as a lab-only capability and agentic RL as something a small team can run. A 14.6 point gain from 3,500 lines of code and an open-weights base is a large jump for a modest setup.

What's still unclear

The evidence here is one paper's reported numbers on one benchmark with one base model. It doesn't say what the training run cost in compute, or whether the gains transfer to a different base model without the same tuning effort. Whether anyone outside the original team reproduces this on a different codebase is the thing worth watching next.