Evidence: benchmark results vendor-reported
DiffusionGemma 26B A4B — VRAM, speed & local setup
Written by Jakub Rusinowski · Last updated
Gemma 4 26B A4B's mixture-of-experts (8 of 128 experts plus 1 shared active per token) turned into a block-diffusion model: an encoder caches the prompt, and a decoder denoises a 256-token canvas in parallel before appending it and starting the next. It needs the same memory as Gemma 4 26B A4B, takes text, images and video, and keeps thinking mode and function calling. Google calls it experimental and reports lower benchmark scores than Gemma 4 26B A4B on most tasks.
DiffusionGemma 26B A4B needs about 16 GB of VRAM at Q4_K_M — quantized weights plus framework overhead, before any KV cache. On Apple Silicon that figure comes out of unified memory.
Get DiffusionGemma 26B A4B running
The cheapest catalogued GPU that runs DiffusionGemma 26B A4B is the AMD Radeon RX 7900 XT (20 GB).
or compare on Vast.ai from $0.35/hr (typical low · varies)
As an Amazon Associate we earn from qualifying purchases. Cloud GPU links are referral links — we may earn a commission at no extra cost to you.
How to run DiffusionGemma 26B A4B
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.
Specifications
Verified — Checked against the primary source — the model card or the vendor spec page — and corroborated by a second independent source.
- Parameters
- 25.2 Billion (3.8B active) + ~0.55B vision encoder
- Context window
- 262,144
- Architecture
- Block-diffusion Mixture-of-Experts (128 experts, 8 active + 1 shared), 25 sliding-window (1,024) + 5 global layers, 256-token canvas
- Provider
- Google DeepMind
- Licence
- Apache-2.0
- Availability
- Weights downloadable
- Practical deployment
- Single consumer GPU
- Specified at
- Q4_K_M
- System RAM
- 32 GB
- Record updated
- 2026-11-21
Commercial use permitted. No usage restrictions beyond attribution.
Limits and caveats
- Decodes by diffusion: speed depends on GPU compute, not memory bandwidth, so this page gives no speed estimate. Google reports up to 4x faster decoding on dedicated GPUs.
- Experimental: Google reports lower scores than Gemma 4 26B A4B on most of its benchmarks.
- Stock llama.cpp, Ollama and LM Studio cannot run it; the GGUF files need the unmerged llama.cpp pull request #24423.
- Benchmarks on the card are vendor-reported.
See also Gemma 4 26B-A4B
Quality and use cases
Scores as published by the model’s authors or an independent evaluator — quality, not throughput, and not measured by us.
DiffusionGemma 26B A4B — frequently asked questions
How much VRAM does DiffusionGemma 26B A4B need?
About 16 GB at Q4_K_M — quantized weights plus framework overhead, before any KV cache. We publish no context-length figure or speed for it: it writes text by diffusion, refining blocks of 256 tokens in parallel, so its speed depends on GPU compute rather than memory bandwidth, which our speed model cannot express. Apple Silicon counts unified memory toward the same figure.
Does DiffusionGemma 26B A4B run on an RTX 4090 (24 GB)?
Its weights and overhead come to about 16 GB at Q4_K_M, but we do not publish a fit verdict or a speed for it: it writes text by diffusion, refining blocks of 256 tokens in parallel, so its speed depends on GPU compute rather than memory bandwidth, which our speed model cannot express.
How do I run DiffusionGemma 26B A4B locally?
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. Running the published tag would send your prompts to a hosted GPU rather than your own machine.