What is DiffusionGemma?
DiffusionGemma is Google's Gemma 4 26B A4B mixture-of-experts model, with about 4B parameters active, retrained to write by diffusion. It fills a 256-token block with placeholder tokens, refines the whole block over several passes, then moves to the next block. That keeps one GPU busy, so a single user gets far more tokens a second.
The price is accuracy. In Google's own table it trails the regular Gemma 4 26B A4B on nearly every test, and Google still recommends Gemma 4 when quality matters most. It is an experimental release with no hosted API from Google, and its knowledge stops at January 2025.
How is DiffusionGemma different from Gemma 4?
Same base model, different way of writing: DiffusionGemma trades some accuracy for much faster output on one GPU.
- Decoding
- Gemma 4 writes one token at a time; DiffusionGemma drafts 256-token blocks and refines each one in parallel.
- Speed
- Up to 4x faster for a single user on a dedicated GPU, with less gain on Apple Silicon or busy servers.
- Quality
- Lower on nearly every benchmark Google published, for example 47.6% against 64.8% on BigBench Extra Hard.
- Size
- The same 26B A4B mixture-of-experts design, fine-tuned to denoise, so it needs about the same memory.
Who is DiffusionGemma best for?
DiffusionGemma is best for developers who want very fast text generation on their own NVIDIA GPU, for in-line editing and code infilling, and for research into text diffusion.
Fast local generation
On one dedicated GPU it writes 700+ tokens a second on a GeForce RTX 5090 and 1,000+ on an H100, which suits chat, drafts and quick code edits for one person.
In-line editing and code infilling
Each block is refined with attention in both directions, so it can fill a gap using the text after it, not only the text before it.
Grid and constraint tasks
Google's Sudoku fine-tuning recipe took it from almost no solved puzzles to 80%, a task that left-to-right models find hard.
Research on text diffusion
The weights are open under Apache 2.0, with training recipes in Google's Hackable Diffusion toolbox and fine-tuning support in Unsloth and NVIDIA NeMo.
What's included with DiffusionGemma?
Block diffusion decoding, a thinking mode, image and video input, function calling, a 256K context, and support in vLLM, Transformers, SGLang and MLX.
- Block diffusion decoding
- Drafts 256 tokens at a time and refines them in parallel, stopping early once the block settles, so simple prompts and code finish in fewer steps.
- Thinking mode
- A switchable reasoning channel lets it work step by step before it writes the final answer.
- Image and video input
- Reads images at variable resolution and video up to 60 seconds at one frame a second, and answers in text.
- Function calling
- Native structured tool use for agent workflows, plus system prompts as in Gemma 4.
- Long context and languages
- Up to 256K tokens of context, with out-of-the-box support for 35+ languages.
- Serving and deployment
- Runs on vLLM, Hugging Face Transformers, SGLang and MLX, or deploys from Google Cloud's Model Garden and as an NVIDIA NIM container.
How much does DiffusionGemma cost?
DiffusionGemma is free: the Apache 2.0 weights download at no cost from Hugging Face and Kaggle, and Google sells no per-token API for it, so you pay only for your own GPU or cloud hardware.
- Download from Hugging Face or Kaggle
- One 26B mixture-of-experts model, about 4B active
- Fits in 18GB of GPU memory when quantized
Google's Gemini API pricing lists Gemma 4 but not DiffusionGemma, so there is no hosted per-token rate from Google. Besides running it on your own GPU, you can deploy it from Google Cloud's Model Garden or as an NVIDIA NIM container and pay for that compute, or serve it with vLLM, SGLang, MLX or Hugging Face Transformers. Full pricing on ai.google.dev. We checked these prices on .
How do readers rate DiffusionGemma?
No one has rated DiffusionGemma yet. Be the first.
Sign in to rate. One rating per account.
DiffusionGemma pros and cons
It is the fastest way to run a Gemma-class model on one GPU, but it is less accurate than Gemma 4, shaky with tool calls and gains little on Macs.
Pros
- Very fast for one person on one NVIDIA card: Google measures 1,000+ tokens a second on an H100, and users report about 670 on an RTX Pro 6000 and over 900 on an RTX 5090.
- Developers on Reddit and Hacker News call the quality usable for its size, especially for latency-bound jobs such as extraction pipelines and quick drafts.
- Apache 2.0 weights, so you can fine-tune it and ship it in commercial products without asking Google.
- Two-way refinement helps where later text constrains earlier text: Google's Sudoku fine-tune went from almost none solved to 80%.
Cons
- Clearly less accurate than Gemma 4 26B A4B: 73.2% against 82.3% on GPQA Diamond and 69.1% against 77.1% on LiveCodeBench v6 in Google's own table, a gap users notice too.
- Tool calls are shaky in practice: users report broken JSON in file-write calls and missing structured output in their setups.
- The speed-up needs a strong dedicated GPU: Google warns Apple Silicon Macs may not see it, one Mac user reports about 15 tokens a second, and it adds little in busy cloud serving.
- Needs about 18GB of GPU memory even quantized, so 16GB cards are out, and mainline llama.cpp does not run it yet.
Is DiffusionGemma open source?
Yes. DiffusionGemma's code is public under the Apache 2.0 licence, so you can run it on your own server.
What are the best DiffusionGemma alternatives?
DiffusionGemma is an open source alternative to Mercury 3 and Mercury 2.5.
Mercury 3FreemiumInception's flagship diffusion language model, writing over 1,000 tokens a second with 1M context.
Mercury 2.5PaidChoose it when you want diffusion speed without running a GPU: Inception hosts it at $0.20 / $0.75 per million input / output tokens, with a 260K context, tool use and structured output.
Gemma 4Open SourceChoose it when answer quality matters more than speed: in Google's own table the same-size 26B A4B reaches a 1718 Codeforces Elo against DiffusionGemma's 1429, and Google serves Gemma 4 free on the Gemini API.
Qwen3.6-27BOpen SourceChoose it when you are building a local coding agent and can accept normal token-by-token speed; it is a dense 27B model under Apache 2.0 whose 262K context stretches to about 1M tokens.
Questions about DiffusionGemma
What is DiffusionGemma?
DiffusionGemma is Google DeepMind's experimental open model that writes text in 256-token blocks instead of one token at a time, for developers who want very fast generation on their own GPU.
Who is DiffusionGemma for?
DiffusionGemma is best for developers who want very fast text generation on their own NVIDIA GPU, for in-line editing and code infilling, and for research into text diffusion.
Is DiffusionGemma free?
Yes. DiffusionGemma's weights are free to download from Hugging Face or Kaggle under the Apache 2.0 license. Google's Gemini API does not list it with a per-token price, so running it yourself costs only the GPU or cloud compute you use.
Is DiffusionGemma open source?
Yes. DiffusionGemma's code is public under the Apache 2.0 licence, so you can run it on your own server.
What are the best DiffusionGemma alternatives?
The closest alternatives on Recaply are Mercury 3, Mercury 2.5, Gemma 4 and Qwen3.6-27B.
Where can I try DiffusionGemma?
On its official website, blog.google.
What hardware does DiffusionGemma need?
DiffusionGemma fits in about 18GB of GPU memory when quantized, so a GeForce RTX 4090 or 5090 can run it. Google tuned it for NVIDIA cards from those up to Hopper and Blackwell servers, and its speed gain depends on a dedicated GPU with plenty of compute.
Does DiffusionGemma run on a Mac?
It runs through MLX, but Google warns that Apple Silicon Macs have less compute relative to memory bandwidth, so they may not see the speedup over Gemma 4 that dedicated GPUs get. On a Mac, standard Gemma 4 may be the better choice.
Can I use DiffusionGemma for commercial work?
Yes. Google releases DiffusionGemma under the Apache 2.0 license, which allows commercial use, changes and redistribution, including fine-tuned versions, as long as you keep the license and copyright notices with what you share.
Is there a DiffusionGemma API?
Not from Google's Gemini API, which lists Gemma 4 but not DiffusionGemma. You can deploy it on Google Cloud from Model Garden, run NVIDIA's NIM container, or serve it yourself with vLLM, which gives you an OpenAI-compatible endpoint.
Our verdict
DiffusionGemma is worth trying when raw speed on your own NVIDIA GPU matters more than getting the best answer.
Google measures up to 4x faster generation than token-by-token models and it runs in 18GB once quantized, but it scores well below Gemma 4 26B A4B on Google's own tests.
- Skip it for
- Agents that need reliable tool calls, Mac-only setups, and any work where accuracy comes first; Gemma 4 fits those better.
- Check first
- Test it on your own GPU, runtime and prompts first, because speed and tool-call reliability vary a lot between setups.
Gemini 3.5 Flash
Nemotron 3.5 Lightning
Bonsai 27B
Step 3.7 Flash
Mercury 2
LFM3