SWE-Next-14B is a repository-level software engineering agent fine-tuned from Qwen/Qwen2.5-Coder-14B-Instruct on the released SWE-Next SFT Trajectories. The model is trained with full-parameter supervised fine-tuning on execution-grounded trajectories collected from real merged pull requests and validated repository environments.
Introduction
SWE-Next introduces reusable repo-quarter profiles, which reuse the same environment across nearby commits in time while keeping each task run separate and reproducible. Using only 30 hours and 639GB of environment storage, SWE-Next processes 3,971 seed repositories and 102,582 candidate commit pairs mined from real merged PRs to construct a dataset of 2,308 self-verifying instances. SWE-Next improves downstream pass@1 on SWE-Bench Verified and SWE-Bench Lite with fewer or comparable training trajectories, making large-scale executable data collection far more practical and accessible for research.
SWE-Next teaser
Model Overview
This model is trained on 3,693 selected SFT trajectories derived from the SWE-Next collection. The training data emphasizes clean repository-level repair traces and recovery-style debugging trajectories rather than isolated code-completion examples.
Training recipe summary:
Base model: Qwen/Qwen2.5-Coder-14B-Instruct
Finetuning: full-parameter SFT
Context length: 32,768
Learning rate: 1e-5
Scheduler: cosine
Dataset: TIGER-Lab/SWE-Next-SFT-Trajectories
Usage
For full usage details, please refer to the official SWE-Next GitHub repository. The repository provides the complete setup and evaluation workflow for released models, including:
environment and dependency installation,
dataset and trajectory downloads,
training configurations for the 7B and 14B models,
vLLM serving commands and repository-level evaluation scripts.
In particular, the GitHub repo contains the exact commands used to serve SWE-Next-14B and evaluate it on SWE-Bench-style tasks under the SWE-Next execution interface.
Relationship to the SWE-Next Release
This repo contains the released 14B model checkpoint. Related artifacts are available separately:
1@misc{liang2026swenextscalablerealworldsoftware,
2 title={SWE-Next: Scalable Real-World Software Engineering Tasks for Agents},
3 author={Jiarong Liang and Zhiheng Lyu and Zijie Liu and Xiangchao Chen and Ping Nie and Kai Zou and Wenhu Chen},
4 year={2026},
5 eprint={2603.20691},
6 archivePrefix={arXiv},
7 primaryClass={cs.SE},
8 url={https://arxiv.org/abs/2603.20691},
9}
