Logics-SWE-Qwen3.6-35B-A3B

Logics-SWE-Qwen3.6-35B-A3B is a software engineering model developed from Qwen3.6-35B-A3B using category-aware multi-teacher on-policy distillation (MOPD). It brings together the capabilities of three software engineering experts in a single model for repository-level coding agents.

Paper · GitHub: AgenticBigBang · Logics-SWE-Qwen3.6-27B · Dataset: Logics-SWE-Env-2.5K

📰 News

• [2026.10.30] Introduced Logics-SWE-Qwen3.6-35B-A3B, trained with MOPD using the same three 27B category experts as Logics-SWE-Qwen3.6-27B.

Highlights

• One policy, three experts. The model integrates three category experts into one student that can work across repository domains without selecting an expert at inference time.
• Reuse of 27B teachers. The same three final Qwen3.6-27B-based experts used for Logics-SWE-Qwen3.6-27B are reused to train the 35B-A3B student.
• Improved SWE and coding results. Compared with Qwen3.6-35B-A3B, the model improves by 3.69 points on SWE-bench Pro V2, 3.00 points on SWE-bench Multilingual, and 2.37 points on SWE-bench Pro.

How It Works

Repository-level software engineering tasks span different domains, from services and data systems to user-facing applications and developer tools. The RRE–MOPD approach develops complementary specialists and combines their knowledge into a single agent model:

1. Organize tasks. SWE Labeler assigns tasks to repository-domain categories: service, data and security; user-facing applications; and systems, tooling and runtimes.
2. Develop experts. Three Qwen3.6-27B-based experts are trained with Refresh–Repair–Expand (RRE), alternating agentic reinforcement learning, mastery refreshes, and supervised repair on verified successful trajectories.
3. Integrate with MOPD. A student initialized from Qwen3.6-35B-A3B generates its own trajectories and receives token-level guidance from the expert corresponding to each training task. The student's initial policy serves as the frozen reference for extrapolation.

The final model runs as a single policy. It does not require the three experts or category routing when deployed.

Experimental Results

The table shows three-round mean scores (%) for the full model-and-agent evaluation setups. Improvements are percentage points over Qwen3.6-35B-A3B.

Hard-51 is a subset of SWE-bench Pro V2. Differences are calculated before rounding the displayed scores.

Intended Use

The model is intended for research and development on repository-level software engineering agents, including codebase exploration, issue diagnosis, code editing, test execution, and iterative debugging. A coding-agent harness can provide the model with repository access, tools, execution feedback, and a workspace in which to build and validate a patch.

Serving

For a local copy of the model, a vLLM server can be started with:

vllm serve /path/to/Logics-SWE-Qwen3.6-35B-A3B

The server exposes an OpenAI-compatible API for integration with a coding-agent harness. Use the model's tokenizer and chat template when constructing requests. See the Qwen model card and vLLM documentation for serving setup.

Data and Resources

The public Logics-SWE-Env-2.5K release provides 2,553 software engineering task instances and matching environment-image references. The AgenticBigBang repository contains SWE Labeler, research artifacts, and evaluation materials.

Citation

If you use this work, please cite:

@misc{zhao2026onetomoremoretoone,
  title         = {One to More, More to One: Category-Aware Iterative Expert Training for Software Engineering Agents},
  author        = {Jie Zhao and Ziyu Jiang and Suhang Zheng and Minghui Shan and Xiaoxiao Xu and Lin Qu},
  year          = {2026},
  eprint        = {2609.23377},
  archivePrefix = {arXiv},
  primaryClass  = {cs.SE},
  url           = {https://arxiv.org/abs/2609.23377}
}

Acknowledgements

We thank the Qwen, ROLL, R2E-Gym, SWE-agent, SWE-bench, SWE-bench Pro, and SWE-bench Multilingual contributors, and the research communities behind MiniRL, MOPD, and ExOPD.

License

Model license: Apache License 2.0.