What is DiffusionGemma 26B A4B?
DiffusionGemma is an experimental open model from Google DeepMind, launched on 10 June 2026. It takes the Gemma 4 26B A4B mixture-of-experts design (25.2 billion parameters, 3.8 billion active, 8 of 128 experts per token) and changes how text is written. Instead of producing one token at a time, it drafts a block of 256 tokens and refines the whole block over several denoising passes, then moves to the next block.
That shifts the bottleneck on a single GPU from memory bandwidth to compute. Google reports more than 1,000 tokens a second on one NVIDIA H100 and more than 700 on a GeForce RTX 5090, and says a quantised build fits in 18 GB of VRAM. It reads text, images and video frames, writes text, has a thinking mode and function calling, and keeps Gemma 4's 256K-token context. It has no audio input.
The weights are under Apache 2.0. In the last 30 days they drew about 636,000 Hugging Face downloads.
Is DiffusionGemma 26B A4B worth it?
DiffusionGemma trades quality for speed, and Google's own card shows how much. Against the standard Gemma 4 26B A4B it scores 77.6% on MMLU Pro against 82.6%, 69.1% on AIME 2026 against 88.3%, 69.1% on LiveCodeBench v6 against 77.1%, and 56.2% on the Tau2 tool-use test against 68.2%. Vision drops harder: 54.3% on MMMU Pro against 73.8%, and twice the OmniDocBench error (0.319 against 0.149). Long-context recall at 128K falls from 44.1% to 32.0%.
What you get back is latency. Producing 15 to 20 tokens per forward pass, it passes 1,000 tokens a second on an H100 and 700 on an RTX 5090, which Google puts at up to four times faster than typical models on dedicated GPUs. Simple prompts and code need fewer denoising steps, so speed varies with the task. Because every token in a block sees the others, it can also fill gaps in code and fix its own mistakes before finishing a block.
The speed only shows up in one setting. Google says the advantage is strongest at low to medium batch sizes on one accelerator, fades under high-traffic cloud serving, where it can cost more to run, and may not appear at all on unified-memory machines such as Apple Silicon Macs. llama.cpp support was announced as coming soon at launch.
That makes it a tool for interactive local work: inline editing, autocomplete, quick drafts where waiting is the bigger cost than a weaker answer. For batch jobs, agents or document reading, the autoregressive Gemma 4 26B is better on every benchmark Google publishes.
Where it earns its place
A developer-tools engineer at an IDE start-up
Offer local inline code edits that appear almost instantly
Serve the quantised model on an RTX 5090 with vLLM and use it for infilling, where its bidirectional blocks help. Google cites 700-plus tokens a second on that card; keep Gemma 4 for harder refactors.
A technical writer at a software company using a workstation GPU
Draft and rewrite documentation sections without waiting on each paragraph
Run it locally with thinking off for quick drafts. Google says it reliably closes complex Markdown; quality is below Gemma 4, so review facts and code samples.
Before you commit
- Setup
- pip install the latest transformers and load DiffusionGemmaForBlockDiffusion, or serve with vLLM or MLX. Google recommends its entropy-bounded sampler with up to 48 denoising steps, usually stopping after 12 to 16.
- Integrations
- transformers, vLLM with Red Hat support, MLX, NVIDIA NIM, Unsloth and NeMo for fine-tuning, and Google's Hackable Diffusion JAX library.
- Data and privacy
- Self-hosted, so prompts stay on your hardware. Apache 2.0 permits commercial use and modification.
- Cost at scale
- Weights are free; one RTX 4090 or 5090 class card runs the quantised build. Google warns that high-traffic cloud serving can cost more than an autoregressive model.
- Limits
- Experimental release; image and video input only, no audio; video up to 60 seconds at one frame a second; thoughts from earlier turns must be stripped from chat history.
What’s good
- Over 1,000 tokens a second on one H100, 700 on an RTX 5090
- Quantised build fits in 18 GB of VRAM
- Apache 2.0 with vision, thinking mode and function calling
- Bidirectional blocks suit code infilling and self-correction
What’s not
- Below Gemma 4 26B on every published quality benchmark
- AIME 2026 falls from 88.3% to 69.1%
- Speed gain fades at high batch sizes and on Macs
- Labelled experimental by Google; no audio input
Best for
Developers building single-user, latency-sensitive tools such as inline editors or autocomplete on one NVIDIA GPU.
DiffusionGemma 26B A4B alternatives
DiffusionGemma 26B A4B: common questions
- Is DiffusionGemma 26B A4B worth it?
- DiffusionGemma makes a capable open model answer one user several times faster on one GPU, and Google itself says to stick with plain Gemma 4 when quality matters. We rate it 3.8 out of 5.
- How much does DiffusionGemma 26B A4B cost?
- Free Apache 2.0 weights. No first-party per-token price published; you pay for your own GPU or cloud deployment.
- What are the downsides of DiffusionGemma 26B A4B?
- Below Gemma 4 26B on every published quality benchmark; aIME 2026 falls from 88.3% to 69.1%; speed gain fades at high batch sizes and on Macs; labelled experimental by Google; no audio input.
- Who is DiffusionGemma 26B A4B best for?
- Developers building single-user, latency-sensitive tools such as inline editors or autocomplete on one NVIDIA GPU.
- How much does DiffusionGemma 26B A4B cost for a team or in production?
- Weights are free; one RTX 4090 or 5090 class card runs the quantised build. Google warns that high-traffic cloud serving can cost more than an autoregressive model.
- Is DiffusionGemma 26B A4B safe for company data?
- Self-hosted, so prompts stay on your hardware. Apache 2.0 permits commercial use and modification.
- What are the best DiffusionGemma 26B A4B alternatives?
- The closest alternatives we have reviewed are Gemma 4, Qwen3.8-27B, gpt-oss-20b and Mistral Small 4.
Answers checked Nov 7, 2026.