Diffusion language models: AI that writes all the words at once

Most chatbots write like a person typing: one word after another, from left to right. Diffusion language models work in a different way. They start with a row of blanks and fill in many words at once, improving the whole text over a few rounds.

They can be much faster, and in 2026 they are a hot research topic. This is how they work.

First: how normal chatbots write

Most of today's chatbots are autoregressive. That means they write one token (a word or a piece of a word) at a time, and each new token depends on all the tokens before it.

This works very well, but it is slow for long answers. Each token has to wait for the one before it.

The idea behind diffusion

Diffusion first became famous for making images. Image tools like Stable Diffusion start with random noise and remove a little of it at each step, until a clear picture appears.

For text, the "noise" is usually hidden words. A diffusion language model starts with every word hidden. At each step, it guesses many hidden words at the same time, keeps the guesses it is sure about, and keeps working on the rest.

Here is a simple picture of the idea, for the sentence "The cat sat on the mat":

• Start: _ _ _ _ _ _
• Step 1: The _ _ on _ mat
• Step 2: The cat _ on the mat
• Step 3: The cat sat on the mat

A normal chatbot needs six steps for six words. Here, three steps were enough. Real answers have hundreds of tokens, so the time saved can be large.

LLaDA: proof that it can work at scale

In February 2025, a team led by Shen Nie shared LLaDA (Large Language Diffusion with mAsking). They trained a diffusion language model with 8 billion parameters from scratch.

They report that LLaDA 8B is competitive with LLaMA3 8B, a normal model of the same size, at learning from examples in the prompt. After extra training on instructions, it could also hold a multi-turn chat. This mattered because many people believed only autoregressive models could work well for language. LLaDA's code is public on GitHub.

Why people care: speed

Mercury Coder, from the company Inception Labs, is a diffusion model for writing code. In independent tests by Artificial Analysis, Mercury Coder Mini produced 1,109 tokens per second and Mercury Coder Small 737 tokens per second, on NVIDIA H100 GPUs. The company reports it is up to 10 times faster than speed-focused normal models, with similar quality.

Google DeepMind has also shown Gemini Diffusion. As of November 2026, Google still calls it an experimental demo.

The weak points

• Words depend on each other. When a model fills many blanks in one step, it guesses each one partly on its own. If two words must fit together, the text can break. Fewer steps means faster, but often worse, text.
• Settings matter a lot. A June 2026 study by Thomas Bertolani and colleagues tested eight diffusion language models on eight benchmarks. Results depended strongly on choices made while generating, such as the number of steps, with a clear trade-off between quality and speed.
• They are new. Most of today's leading chatbots are still autoregressive.

A big research area in 2026

NeurIPS 2026 has a full workshop called "Diffusion Language Models: Foundations, Efficiency, and Reasoning" in December. Open questions include how to make these models reason well, how to train them cheaply and how to choose the number of steps. That means there is room for new researchers, including students.

How to start as a student

1. 1Learn how image diffusion works first. It is the easiest way to get the core idea.
2. 2Read the LLaDA paper (arXiv:2502.09992) and look at its public code.
3. 3Try a small experiment: change the number of steps and measure how speed and quality change.
4. 4Read the 2026 study by Bertolani and colleagues to see how to compare models fairly.

New to research papers? Start with how to read a research paper. At APJ-RAISE, students pick up topics like this and work on them together in small teams. Joining is free.