metadata
base_model: Qwen/Qwen3-8B
language:
- en
license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
tags:
- speculative-decoding
- specblock
- draft-model
SpecBlock-Qwen3-8B
SpecBlock draft model for speculative decoding, trained against the target model Qwen/Qwen3-8B.
This model was introduced in the paper SpecBlock: Block-Iterative Speculative Decoding with Dynamic Tree Drafting.
Method
SpecBlock — multi-block test-time training with cross-slot hidden injection between decoder layers and dynamic tree drafting.
Usage
End-to-end training and inference code can be found in the official repository: https://github.com/shiweijiezero/SpecBlock
Quick eval with the HF backend:
python benchmarks_hf/run_eval.py \
--algorithm specblock \
--model-path Qwen/Qwen3-8B \
--draft-model-path <local-clone-of-this-repo> \
--benchmark-list mtbench:80 humaneval:164 gsm8k:200 \
--output ./hf_results/specblock_qwen3.jsonl
Citation
@misc{shi2026specblockblockiterativespeculativedecoding,
title={SpecBlock: Block-Iterative Speculative Decoding with Dynamic Tree Drafting},
author={Weijie Shi and Qiang Xu and Fan Deng and Yaguang Wu and Jiarun Liu and Yehong Xu and Hao Chen and Jia Zhu and Jiajie Xu and Xiangjun Huang and Jian Yang and Xiaofang Zhou},
year={2026},
eprint={2605.07243},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2605.07243}
}