Diffusion Large Language Models
summary
Inception builds a diffusion LLM called Mercury 2 that claims to be the fastest model on the Artificial Analysis index and to deliver 3–10x latency speedup versus frontier autoregressive models while matching quality. Training uses a mean-field log-likelihood estimator to make GRPO-style on-policy RL tractable (reducing variance and memory cost), and the company reports 3–5x faster post-training scaling in practice; deployments include voice, search and coding where latency and cost drop noticeably.
AI-written summary · about
chapters (7)
- 0:00Introduction to diffusion LLMs
- 2:56How diffusion LLMs work
- 8:35Evolution of diffusion LLMs
- 10:21Modern applications of diffusion LLMs
- 17:10Challenges in training diffusion LLMs
- 24:07Why diffusion for universal AI
- 27:40Q&A: Scaling laws and inference
related talks
45mDiffusion to Prod: vLLM-Omni Unifies Multimodal Inference
14mScalable Solutions for Running Large Language Models
16mHow fast are LLM inference engines anyway?
28mText Diffusion
39mAccelerating Open-Source RL and Agentic Inference with vLLM
01h 02mAdvancing Diffusion Models for Text Generation
24mLLM Inference Optimization From Simple To Sophisticated
41mSub-Linear Algorithms meets Large Language Models