LLM Hacker News 6h ago 2 min read
Hacker News Zeroes In on I-DLM as a Diffusion LLM That Might Keep AR Quality Without Giving Up Speed
Hacker News readers are treating this less like another diffusion-text curiosity and more like a possible faster serving path that still stays close to autoregressive quality. The project page claims I-DLM-8B reaches 69.6 on AIME-24, 45.7 on LiveCodeBench-v6, and 2.9-4.1x higher throughput at high concurrency.