Diffusion language models
Models that write by refining a whole passage at once, such as Mercury and Gemini Diffusion. 5 models on EveryLLM so far.
Diffusion language models write differently from the usual next-token models. They start from a passage that is fully masked or noisy and refine all of it over a series of steps, filling in many tokens at once, in the way image generators turn noise into a picture. Because they are not bound to one token at a time, they can be much faster, and they can revise earlier parts of an answer. The approach is still young.
Research roots
Diffusion models transformed image generation from 2020. Adapting them to text, whose tokens are separate symbols rather than continuous pixel values, took new methods: Diffusion-LM from Stanford (2022), and score entropy discrete diffusion, which won a best paper award at the ICML conference in 2024. LLaDA, an 8-billion-parameter model from Renmin University of China and Ant Group released in February 2025, showed a diffusion model competing with conventional models of similar size.
First products
Inception Labs, a start-up founded by researchers from Stanford, UCLA and Cornell, released Mercury Coder in 2025 and reported speeds above 1,000 tokens per second on an Nvidia H100 chip for its smaller version. Google DeepMind showed an experimental Gemini Diffusion model at Google I/O in May 2025. Diffusion models still trail the best conventional models on reasoning, and the software for serving them is less mature.