ByteDance tests diffusion code model

ByteDance’s Seed team has introduced Seed Diffusion Preview, a discrete diffusion language model built for structured code tasks. Processing at 2,146 tokens per second, it is said to deliver 5.4 times faster inference than comparable autoregressive models while maintaining similar benchmark accuracy. The model combines edit-based training, constrained decoding, and on-policy learning to advance diffusion toward scalable real-world applications.