ByteDance releases experimental diffusion language model Seed Diffusion

The Seed team at ByteDance has released an experimental diffusion language model called Seed Diffusion Preview. This model validates the feasibility of discrete diffusion technology as the foundational framework for next-generation language models. By leveraging key technologies such as two-stage diffusion training, constrained sequence learning, and efficient parallel decoding with reinforcement, it achieves an inference speed of 2,146 tokens per second—5.4 times faster than autoregressive models of comparable scale.