What if an AI could write a whole paragraph at once, instead of typing it out one word at a time? That is the big idea behind Google’s newest model, and it makes text generation up to four times faster.

Google just released DiffusionGemma, an experimental open AI model that takes a clever new approach to generating text. In this guide we will explain what it is and why it is so fast. We will also cover how it works, all in plain words anyone can follow. Let us get into it.

Fast Facts

What Is DiffusionGemma

Let us start simple. As Google shared in its official announcement, DiffusionGemma is an experimental open AI model that generates text in a brand new way, much faster than usual.

It comes from the same Gemma family as Google’s other open models, but it borrows an idea from AI image generators. Instead of writing word by word, it shapes a whole block of text at once.

It is free to download under an Apache 2.0 license. It is aimed at developers and researchers who want speed for local, interactive apps.

Quick Take: DiffusionGemma is not meant to replace Google's top-quality models. Think of it as a fast, experimental tool built for speed-first jobs, like live editing or rapid prototyping.

Why Is It So Much Faster

Here is the heart of it, explained with a simple picture. Most AI models work like a typewriter. They produce text one word at a time, left to right, each word waiting for the one before it.

That is fine on big cloud servers, which juggle thousands of users at once. But on your own computer, serving just you, the graphics card spends most of its time waiting for the next keystroke. A lot of power sits idle.

DiffusionGemma flips this around. It drafts a whole 256-word block in one go, giving your hardware a big chunk of work all at once. Google sums it up with a great analogy.

The Simple Version: A normal AI model is like a typewriter, stamping one letter at a time. DiffusionGemma is like a printing press, stamping a whole page at once. Same goal, far less waiting.

How Text Diffusion Works, Step by Step

So how does it write a whole paragraph at once? It uses the same trick as AI image tools, which start with random noise and slowly sharpen it into a picture. DiffusionGemma does this with words.

Here is the process in three easy steps.

1. The blank canvas: The model starts with a block of random placeholder words, a rough draft of pure noise.
2. Refine in passes: It reviews the whole block again and again. It locks in the words that fit, then uses them as clues to fix the rest.
3. Final polish: After a few passes, the random mess sharpens into clean, finished text.

Because the model can see the whole paragraph as it writes, it can do clever things. For example, it can neatly close complex formatting, or write and show code almost in real time.

Why It Matters: Seeing the whole block at once also lets the model fix its own mistakes as it goes. It can spot an error in one spot and correct it using the rest of the text as context.

The Speed and Hardware Numbers

Talk of speed is nice, but here are the real figures Google shared. They show just how quick this model is on the right hardware.

It is also surprisingly light to run. Even though it is a 26 billion parameter model, it only switches on 3.8 billion of them at a time. This design is called Mixture of Experts. When compressed, it fits inside 18GB of memory, which high-end consumer graphics cards have.

Worth Knowing: Mixture of Experts means the model is big in total but only uses the part it needs for each task. You get the smarts of a large model without paying the full speed and memory cost.

What Is It Good For

DiffusionGemma shines at tasks where speed matters and where the whole answer benefits from being seen at once. Some examples Google highlights.

• In-line editing, where you tweak text live and want instant results.
• Code infilling, filling in missing code in the middle of a file, not just at the end.
• Rapid iteration and prototyping, where quick drafts beat perfect ones.
• Puzzle-like and structured tasks, such as Sudoku or math, where answers depend on each other.

In one fun demo, the team Unsloth fine-tuned DiffusionGemma to solve Sudoku. Normal left-to-right models struggle with this, because in Sudoku each box depends on the others. Seeing the whole grid at once is a natural fit.

The Honest Catch You Should Know

Now for the fair print, because this is an experimental model with real trade-offs. It is important to know them before you use it.

• Lower quality: To get its speed, DiffusionGemma gives up some output quality. For top-quality work, Google recommends its standard Gemma 4 model instead.
• Local, not cloud: The speed boost is for running on your own machine for one user. On busy cloud servers, normal models are still better and cheaper.
• Apple Silicon: Macs with Apple chips may not see the same speed gain, due to how their memory works.

Heads Up: Pick the right tool for the job. Choose DiffusionGemma when speed and local interactivity matter most, and stick with standard Gemma 4 when you need the highest quality output.

How to Get Started Today

Want to try it yourself? It is open and available right now. Here is the quick path for developers.

1. Download the weights: Grab the experimental model free on Hugging Face, under the open Apache 2.0 license.
2. Read the guide: Follow Google’s DiffusionGemma developer guide to set it up and learn the basics.
3. Use your favorite tools: It works with popular tools like MLX, vLLM, and Hugging Face Transformers, with support for llama.cpp coming soon.
4. Run it your way: Use your own desktop GPU, or run it in the cloud through Google’s Model Garden or NVIDIA NIM.

Google also worked with NVIDIA to tune it for their chips. So it runs well on home cards like the RTX 5090 and 4090, and on enterprise systems. You can read the full DiffusionGemma developer guide to go deeper.

Frequently Asked Questions

What Is DiffusionGemma

DiffusionGemma is an experimental open AI model from Google that generates text in a new, faster way. Instead of writing one word at a time, it produces whole blocks of text at once, reaching up to 4 times faster speeds on GPUs. It is free to download under an Apache 2.0 license.

How Is DiffusionGemma So Fast

It drafts a full 256-word block of text in parallel, rather than word by word. This keeps your graphics card fully busy instead of waiting between words. Google compares it to switching from a typewriter to a printing press that stamps a whole page at once.

Is DiffusionGemma Better Than Gemma 4

Not for quality. DiffusionGemma trades some output quality for speed, so for the best quality Google recommends standard Gemma 4. DiffusionGemma is designed for speed-first, local, interactive tasks like in-line editing and rapid prototyping.

What Does Text Diffusion Mean

Text diffusion borrows from AI image generators. The model starts with a block of random placeholder words. It then makes several passes to lock in correct words and refine the rest. The text gradually sharpens into a clean final result, much like a photo coming into focus.

What Hardware Does DiffusionGemma Need

It runs best on dedicated GPUs. It reaches over 1000 words per second on a single NVIDIA H100 and over 700 on an RTX 5090. When compressed, it fits within 18GB of memory, so high-end consumer graphics cards can run it locally.

Is DiffusionGemma Free to Use

Yes. Google released DiffusionGemma as an open model under an Apache 2.0 license, so you can download the weights for free from Hugging Face. It works with common developer tools, and you can also run it in the cloud through Google’s Model Garden.

What Can You Build With DiffusionGemma

It suits fast, interactive, local tasks. Good fits include in-line text editing, code infilling, and rapid prototyping. It also handles structured puzzles like Sudoku or math well, where the answer benefits from the model seeing the whole block at once rather than word by word.

Does DiffusionGemma Work on a Mac

It can run, but Macs with Apple Silicon may not see the same speed boost. The speedup relies on heavy parallel computing. Apple’s unified memory design works differently, so the gain over normal models can be smaller.