bingyang-lei/Qwen3-4B-Thinking-Draft-OPD

Qwen3-4B-Thinking-Draft-OPD

This repository contains Qwen3-4B-Thinking-Draft-OPD, a draft model for speculative decoding introduced in the paper Draft-OPD: On-Policy Distillation for Speculative Draft Models.

• —Project Page: https://www.haodilei.top/draft-opd/
• —GitHub Repository: https://github.com/bingyang-lei/Draft-OPD

Model Details

• —Target Model: `Qwen3-4B(enable_thinking=True)`
• —Model type: Draft model for speculative decoding
• —Architecture: Same as the original DFlash draft model
• —Post-training method: Draft-OPD (On-Policy Distillation)

Performance and Training Method

Draft-OPD trains speculative draft models with on-policy target feedback. Instead of only learning from fixed target-generated trajectories, the drafter is supervised on draft-induced states exposed during speculative verification, including the positions where draft proposals are rejected.

Experiments show that Draft-OPD achieves over $5\times$ lossless acceleration for thinking models across diverse tasks, improving over EAGLE-3 and DFlash by 23% and 13%. For detailed training procedures, evaluation settings, and performance results, please refer to our paper: https://arxiv.org/abs/2605.29343.

Citation

If you find our work useful, please consider citing our paper:

@misc{lei2026draftopdonpolicydistillationspeculative,
      title={Draft-OPD: On-Policy Distillation for Speculative Draft Models}, 
      author={Haodi Lei and Yafy Li and Haoran Zhang and Shunkai Zhang and Qianjia Cheng and Xiaoye Qu and Ganqu Cui and Bowen Zhou and Ning Ding and Yun Luo and Yu Cheng},
      year={2026},
      eprint={2605.29343},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2605.29343}, 
}