Stable-DiffCoder
8B open diffusion code LLM that beats comparable autoregressive and diffusion code models, using block-diffusion continued pretraining on the Seed-Coder pipeline.
A controlled test that diffusion-style training can improve code modeling beyond autoregressive training at fixed data and architecture, with a block-wise clipped noise schedule. MIT licence.
- Date
- Thursday, 15 January 2026
- Lab
- ByteDance Seed
- Kind
- open-weights
- Access
- open weights
Date from HF weights upload (2026-01-15 UTC); no dated launch post found. Authors state it surpasses ~8B AR and diffusion code models; no numbers captured.
Sources
- bytedance-seed.github.io/Stable-DiffCoder/
- huggingface.co/ByteDance-Seed/Stable-DiffCoder-8B-Instruct
This record was checked against its sources on 6 October 2026. How we check