Google DeepMind · 2026-06-10 · major
Google Ships DiffusionGemma — Apache-2.0 26B/3.8B-Active Mixture-of-Experts That Denoises 256 Tokens in Parallel via Discrete Block Diffusion, Hits 1,000+ Tokens/Sec on H100 and 700+ on RTX 5090 While Posting 77.6% MMLU Pro, 73.2% GPQA Diamond, and 69.1% LiveCodeBench v6
Open-weight Gemma variant swaps next-token autoregression for discrete diffusion. Iteratively denoises 256-token canvases, appends each to a KV cache, and lands open weights on Hugging Face with Apache-2.0.
Google opens a 26B-parameter Gemma that denoises 256 tokens at once instead of generating them one by one.
Key specs
| Active params | 3.8B |
|---|---|
| MMLU | 77.6% |
| GPQA | 73.2% |
| Total parameters | 25.2B |
| Canvas length | 256 tokens |
| Throughput h100 | 1,000+ tokens/sec |
| Throughput rtx5090 | 700+ tokens/sec |
| Vram quantized | 18GB |
| Live code bench v6 | 69.1% |
| Aime2026 no tools | 69.1% |
| Mmmu pro vision | 54.3% |
What is it?
DiffusionGemma is a new open-weight Gemma model from Google DeepMind that ships under Apache-2.0 on Hugging Face. It is a 25.2B mixture-of-experts with 3.8B parameters active per step, exposed through Transformers, vLLM, SGLang, and Docker Model Runner. The headline is the decoding stack: instead of producing one token at a time, the model fills in a 256-token canvas by iteratively denoising it.
How does it work?
Generation runs as discrete diffusion. The model starts from a noised 256-token canvas, repeatedly denoises it with a diffusion sampler, then runs the finished canvas through the encoder and appends it to the KV cache before starting the next block. That lets the GPU work on many tokens in parallel per forward pass; Google reports 1,000+ tokens/sec on an H100 and 700+ on an RTX 5090, hitting 18 GB of VRAM once quantized. The companion `google/hackable_diffusion` Jax toolbox publishes the training and sampling primitives the team used while building it.
Why does it matter?
Block diffusion is one of the few credible attacks on autoregressive throughput at frontier scale. A 26B MoE that posts 77.6% MMLU Pro, 73.2% GPQA Diamond, and 69.1% LiveCodeBench v6 while sustaining four-figure tokens/sec on a single GPU makes very fast local inference and high-throughput batch serving look practical. Apache-2.0 weights also give labs and tooling vendors a real codebase to study diffusion-style LLM decoding instead of relying on closed APIs.
Who is it for?
Inference-platform builders, agent and code-tool authors chasing throughput, ML researchers studying non-autoregressive LLMs
Try it
huggingface.co/google/diffusiongemma-26B-A4B-it