Diffusion models conquered image generation. Now they're coming for text. A new comprehensive guide from the Kuleshov Research Group lays out the full technical stack behind diffusion language models — the alternative to autoregressive (left-to-right) generation that's gaining serious momentum in 2026.
Unlike standard LLMs that produce text one token at a time, diffusion language models generate the entire sequence at once, starting from random noise and iteratively refining it. This unlocks key advantages: built-in error correction (mistakes can be revised mid-generation), bidirectional context (every position attends to the full sequence), and potentially 5-10x faster inference via parallel generation.
The guide traces the evolution from simple masked diffusion (essentially a generative BERT) through modern production systems. Google's Gemma Diffusion, NVIDIA's Nemotron Diffusion, and Inception Labs' Mercury 2 all ship diffusion-based LLMs today. Mercury 2 achieves over 1,000 tokens/second on standard GPUs — speeds previously possible only with specialized inference chips.
The authors argue diffusion may unlock the next scaling leap: just as transformers unlocked pre-training scaling by parallelizing training, diffusion could unlock inference-time scaling by parallelizing generation.