Google releases DiffusionGemma: an open 26B MoE text-diffusion model with up to 4x faster generation

On June 10, 2026, Google released DiffusionGemma, an experimental open model that uses text diffusion instead of conventional token-by-token generation. Google claims up to 4x faster text generation on GPUs, and the model ships under the Apache 2.0 license.

How it works — blocks of text instead of one token at a time

Typical LLMs generate text like a typewriter, one token at a time. DiffusionGemma applies the diffusion approach known from image generation to text: it generates 256 tokens in parallel per forward pass and refines the output over multiple passes, locking in correct tokens as it goes.

• Architecture: 26B Mixture of Experts (4B active), based on the Gemma 4 architecture (model: diffusiongemma-26B-A4B-it)
• Generation: text diffusion (256 tokens in parallel, iterative refinement)
• License: Apache 2.0 (open, commercial use allowed)

Speed and hardware — 1,000+ tokens/sec on an H100

Google reports 1,000+ tokens per second on a single NVIDIA H100 and 700+ tokens per second on a consumer GeForce RTX 5090. When quantized, the model fits within the 18GB VRAM of high-end consumer GPUs.

• Where to get it: Hugging Face (weights), Google Cloud Model Garden, NVIDIA NIM
• Tooling: MLX, vLLM, Hugging Face Transformers, Unsloth

Use cases and limits — local-first, quality below Gemma 4

The intended use is speed-critical local work such as in-line editing and rapid iteration. Google is explicit about the trade-offs: the model is experimental, overall output quality is lower than standard Gemma 4, and the speedup targets local or low-concurrency inference rather than high-QPS cloud serving.

For anyone running open models on their own GPU, this adds a notably fast option where responsiveness matters more than peak quality.