Seed Diffusion Preview
Discrete-diffusion code model generating at 2,146 tokens/s on H20 GPUs, claimed faster than Mercury Coder and Gemini Diffusion at similar quality.
Parallel, non-sequential decoding avoids token-by-token latency; ByteDance claims a new speed-quality Pareto frontier for code. A preview demo, not an open release; a research follow-up, Stable-DiffCoder, opened 8B weights in January 2026.
- Date
- Monday, 4 August 2025
- Lab
- ByteDance Seed
- Kind
- model
- Access
- research preview
Figures
| Measure | Value | Measured by |
|---|---|---|
| Inference speed on H20 GPUs | 2,146 tokens/s | company |
Date is arXiv v1; the blog demo is believed to be a few days earlier but I could not open its dated text.
Sources
This record was checked against its sources on 6 October 2026. How we check