Diffusion language models, explained simply

In plain words

Most chatbots write one word after another, left to right, like typing. Each word waits for the one before it. That is next-token prediction.

A diffusion language model works differently. It starts with a full-length draft that is all blanks. In the first pass it fills in the words it is most sure about, anywhere in the passage. In the next pass it fills more. Some models can also change earlier guesses. After a handful of passes, the text is complete.

The idea comes from diffusion image models, which turn noise into a picture step by step. Because many words are written at the same time, these models can be much faster. Some can also go back and fix a word they guessed earlier.

As of November 2026, they are mostly experimental or specialised. The main chatbots still write left to right.

Everyday example

Filling in a crossword

Nobody solves a crossword strictly from 1 Across to the last clue. You fill the easy answers first, wherever they are, and each one gives letters that help with the harder ones. Sometimes you rub one out and fix it. A diffusion language model writes text this way: easy words first, then the rest, until the grid is full.

Where the example breaks: A crossword has one correct answer per square. Text has many good wordings, so the model is choosing likely words, not checking against an answer key.

How it actually works

Diffusion for text had to be adapted, because words are separate symbols, not smooth pixel values. The common approach uses masking instead of visual noise.

Training: hide and guess

Take real text and hide a random share of its tokens, anywhere from a few to all of them. A transformer learns to guess the hidden tokens from everything visible, on both sides. It is like BERT's fill-in-the-blank game, played at every level of difficulty.

Writing: unmask over several passes

To write, start with a block of masked slots, for example 256. Each pass predicts every slot, keeps the most confident guesses and re-masks the rest. Fewer passes means faster but rougher text.

Long answers in blocks

Some systems write one block at a time, left to right, and fill each block by diffusion. That mixes the two styles and lets them reuse the KV cache trick for speed.

The trade-offs

They can produce over a thousand tokens per second on a single data-centre GPU, and they can fill gaps in the middle of a text naturally. But quality can lag: Google says its DiffusionGemma, for example, gives lower overall quality than its regular Gemma 4 models.

LLaDA

February 2025: an 8 billion parameter diffusion model from Renmin University of China and Ant Group, competitive with Meta's Llama 3 8B on several tests

Mercury

Inception Labs' commercial diffusion models, first for coding. Reported at over 1,000 tokens per second on an NVIDIA H100

Gemini Diffusion

Google DeepMind's experimental text diffusion model, shown in May 2025

DiffusionGemma

June 2026: Google's experimental open-weight diffusion model, which Google says is up to 4 times faster on GPUs

Questions people ask

Are diffusion language models faster?

Often, yes. Because they write many tokens per pass, published figures run above 1,000 tokens per second on one high-end GPU. The speed depends on how many passes you allow, and fewer passes can mean lower quality.

Is ChatGPT a diffusion model?

No. ChatGPT's models write text one token at a time, left to right. Image tools like Stable Diffusion use diffusion for pictures. Diffusion for text is a newer, mostly experimental approach.

Will diffusion replace next-token prediction?

No one knows yet. Diffusion models are fast and good at filling gaps, but as of November 2026 the leading chatbots still write left to right, and most diffusion text models are labelled experimental. Some research mixes the two, writing in blocks.

How is a diffusion language model different from a diffusion image model?

Image diffusion adds and removes smooth visual noise from pixels. Text is made of separate tokens, so most text diffusion hides tokens behind a mask and learns to reveal them. The step-by-step refining idea is the same.