DiffusionGemma — local AI model by Google DeepMind
Written by Jakub Rusinowski · Last updated
Google's experimental text-diffusion model: the Gemma 4 26B A4B mixture-of-experts reworked to write blocks of 256 tokens in parallel and refine them, instead of one token at a time. It takes text, images and video, supports a 256K context and a thinking mode, and ships under Apache 2.0. Google reports much faster decoding on data-centre GPUs and lower quality than the standard Gemma 4, so it is a speed experiment, not a drop-in replacement. Local GGUF use still needs an unmerged llama.cpp pull-request build.
Variants
The smallest DiffusionGemma variant needs about 16 GB of VRAM at Q4_K_M — quantized weights plus framework overhead, before any KV cache.
| Model | VRAM at Q4 | VRAM | Context | Run it |
|---|---|---|---|---|
| DiffusionGemma 26B A4B → 25.2B (3.8B active) | ~16.4 GB | 262,144 | cloud-hosted tag |
Memory is quantized weights plus overhead at Q4_K_M, from the same engine as the GPU & VRAM checker.
How to run DiffusionGemma locally
Install Ollama, then pull the tag.
No Ollama or LM Studio route: neither can run the diffusion decoder. Official routes are Hugging Face Transformers and vLLM (NVIDIA's NVFP4 build). Local GGUF files need an experimental llama.cpp build from pull request #24423 with llama-diffusion-cli, which was not merged on 21 November 2026.
Pick a size above for its own VRAM figure, speed estimate and install command.
Licence
Commercial use permitted. No usage restrictions beyond attribution.
Applies to: DiffusionGemma 26B A4BRecommended GPU
The cheapest catalogued GPU that runs DiffusionGemma locally (min 16 GB VRAM) is the AMD Radeon RX 7900 XT (20 GB).
DiffusionGemma — frequently asked questions
How much VRAM does DiffusionGemma need?
DiffusionGemma needs about 16 GB VRAM at Q4_K_M quantization for its smallest variant. Variants: DiffusionGemma 26B A4B (16 GB, Q4_K_M). On Apple Silicon, unified memory counts toward this requirement.
Can I run DiffusionGemma on an RTX 4090 (24 GB)?
Yes — DiffusionGemma runs on an RTX 4090 (24 GB) and other 24 GB cards such as the RTX 3090. Smaller variants also fit comfortably on 8–16 GB GPUs at Q4_K_M.
What quantization should I use for DiffusionGemma?
Q4_K_M is the best balance of quality and VRAM for DiffusionGemma in most cases. Choose Q8_0 for near-lossless quality if you have spare VRAM, or smaller quants (Q3/Q2) only when memory is tight.
How do I run DiffusionGemma with Ollama?
DiffusionGemma has no Ollama tag yet. No Ollama or LM Studio route: neither can run the diffusion decoder. Official routes are Hugging Face Transformers and vLLM (NVIDIA's NVFP4 build). Local GGUF files need an experimental llama.cpp build from pull request #24423 with llama-diffusion-cli, which was not merged on 21 November 2026.