LiteCoder-Terminal-30b-a3b-sft

LiteCoder-Terminal-30b-a3b-sft

Paper | Code | Blog Post

LiteCoder-Terminal-30b-a3b-sft is part of our latest release on lightweight code agents. The model is fine-tuned from Qwen3-30B-A3B-Instruct-2507 on the LiteCoder-Terminal-SFT dataset.

Compared to our previous preview version, we scaled up the training data from under 1,000 samples to 11,255 trajectories, incorporating a broader task taxonomy and diverse agent scaffolds. With these updates, the model shows consistent improvements across Terminal Bench evaluations.

Released Artifacts

Results

Terminal Bench 1.0 Performance

Terminal Bench 2.0 Performance

Terminal Bench Pro Performance

Citation

@article{peng2026litecoderterminal,
  title={LiteCoder-Terminal: Scaling Long-Horizon Terminal Environments for Learning Language Agents},
  author={Peng, Xiaoxuan and Zhang, Kaiqi and Lu, Xinyu and Cao, Boxi and Lu, Yaojie and Lin, Hongyu and Han, Xianpei and Sun, Le},
  journal={arXiv preprint arXiv:2605.29559},
  year={2026}
}