LiteCoder-Terminal-4b-sft is part of our latest release on lightweight code agents. The model is fine-tuned from Qwen3-4B-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.
@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}}