SYSTEM NOTICE

Auto translation by AI. Be sure, accuracy, nuances and authorial intent may not be fully reflected.

What is DiffusionGemma? - Text Generation via Diffusion Models

2026年7月19日 21:37

In June 2026, Google announced DiffusionGemma. It is an experimental open model that generates text using a "diffusion" mechanism, and its weights have been released under the Apache 2.0 license.

Compared to traditional LLMs, the biggest feature of DiffusionGemma is that it is four times faster. I found the following article interesting, so I have summarized it.

1. How text generation has worked until now

Like ChatGPT, most large language models generate text one token at a time from left to right. Because they have to wait for the next word after generating each one, there has been a problem with very poor time efficiency.

2. What has become possible with DiffusionGemma

Because DiffusionGemma is based on a diffusion model, it drafts the entire token sequence at once rather than predicting words in order.

Up to 4x faster.
Hardware efficient. Although it is a 26B MoE, only 3.8B are active during inference, and if quantized, it fits on a consumer GPU with 18GB of VRAM.
Self-correction. It can correct errors in real-time while looking at the entire block.

While maintaining intelligence equivalent to other Gemma models, the latency is dramatically improved.

3. How text generation via diffusion models works

You can think of the mechanism as being similar to diffusion models for image generation.
Because it can generate while looking at the whole picture, it can close complex markdown without breaking it, or generate and render code in near real-time.

Prepare a random string of characters
Determine the correct characters through multiple processing steps, and use those as clues to refine the rest
Finally converge to a high-quality output

An example was also shown where it became possible to solve "Sudoku," which traditional LLMs are not good at, through fine-tuning.

4. Use cases and areas where it is not suitable

Quality: As a result of prioritizing speed and parallel generation, the output quality is lower than the standard Gemma 4.

Execution environment: The benefits of speed are limited to local, low-parallel inference. Autoregressive models are more efficient for high-load cloud serving.

Target users: For researchers and developers exploring interactive applications where speed is critical, such as inline editing, fast iterative generation, and code completion.

For a detailed explanation of diffusion models, I recommend Mr. Okanohara's book (this is an affiliate link).