# Seed Diffusion: A Large-Scale Diffusion Language Model with High-Speed Inference

**Authors:** Yuxuan Song, Zheng Zhang, Cheng Luo, Pengyang Gao, Fan Xia, Hao Luo, Zheng Li, Yuehang Yang, Hongli Yu, Xingwei Qu, Yuwei Fu, Jing Su, Ge Zhang, Wenhao Huang, Mingxuan Wang, Lin Yan, Xiaoying Jia, Jingjing Liu, Wei-Ying Ma, Ya-Qin Zhang  
**arXiv:** [2508.02193](https://arxiv.org/abs/2508.02193)  
**Date:** 2025-08-04  
**Summary page:** https://www.summarizepaper.com/en/paper/2508.02193v1/

## Key points

- Utilizes discrete-state diffusion for unparalleled inference speed
- Employs non-sequential, parallel generation for significant speedup compared to traditional methods
- Successfully demonstrated in models like Mercury Coder and Gemini Diffusion
- Achieves impressive inference speed of 2,146 token/s on H20 GPUs
- Maintains competitive results across various standard code evaluation benchmarks
- Surpasses contemporary models like Mercury and Gemini Diffusion in speed
- Team behind the model is constantly pushing the boundaries of language modeling technology

## Summary

is a groundbreaking language model that utilizes discrete-state diffusion to achieve unparalleled inference speed. By employing non-sequential, parallel generation,  offers a significant speedup compared to traditional token-by-token decoding methods. This approach has been successfully demonstrated in recent models such as Mercury Coder and Gemini Diffusion. In terms of performance,  achieves an impressive inference speed of 2,146 token/s on H20 GPUs while maintaining competitive results across various standard code evaluation benchmarks. This speed surpasses contemporary models like Mercury and Gemini Diffusion, positioning  at the forefront of the speed-quality Pareto frontier for code models. The team behind , including Yuxuan Song, Zheng Zhang, Cheng Luo, Pengyang Gao, Fan Xia, Hao Luo, Zheng Li, Yuehang Yang, Hongli Yu, Xingwei Qu, Yuwei Fu, Jing Su, Ge Zhang, Wenhao Huang,Mingxuan Wang,Lin Yan,Xiaoying Jia,Jingjing Liu ,Wei-Ying Ma,Ya-Qin Zhang,Yonghui Wu,Hao Zhou are constantly pushing the boundaries of language modeling technology. For those interested in exploring  further or trying out a demo of the model's capabilities can visit https://studio.seed.ai/exp/seed_diffusion/. Additionally more information about the project can be found at https://seed.bytedance.com/seed_diffusion.