字节跳动Seed团队发布扩散语言模型，每秒推理速度2146 tokens

Core Insights - ByteDance's Seed team released an experimental diffusion language model called Seed Diffusion Preview, aiming to validate the feasibility of discrete diffusion technology as a foundational framework for next-generation language models [1] Group 1: Model Performance - The experimental results indicate that the code inference speed of Seed Diffusion Preview can reach 2146 tokens per second, which represents a 5.4 times improvement in speed compared to autoregressive models of similar scale [1]