Google announced DiffusionGemma on June 10, 2026, an experimental open model that rethinks how text is generated. Released under an Apache 2.0 license, the 26-billion-parameter Mixture-of-Experts (MoE) model moves away from the sequential, token-by-token approach of conventional large language models. Instead, it produces entire blocks of text simultaneously, which the company said delivers up to 4x faster generation on dedicated GPUs. Research scientists Brendan O'Donoghue and Sebastian Flennerhag said the model is built on the Gemma 4 family and draws on Google's Gemini Diffusion research.
According to figures shared by Google, DiffusionGemma generates more than 1,000 tokens per second on a single NVIDIA H100 and more than 700 tokens per second on an NVIDIA GeForce RTX 5090. Although the model holds 26 billion total parameters, only 3.8 billion are active during inference, and when quantized it fits within 18GB of VRAM, allowing it to run on high-end consumer GPUs. The model generates 256 tokens in parallel on each forward pass, and its bi-directional attention lets every token attend to all others. Google said this offers advantages for non-linear tasks such as in-line editing, code infilling and mathematical structures.
The way DiffusionGemma works resembles AI image generators. The model begins with a canvas of random placeholder tokens, then makes multiple passes, locking in correct tokens and using them as context clues to refine the rest. While the research community has explored diffusion-based text generation for years, applying it to large models has remained difficult. Google noted that autoregressive models run efficiently in the cloud by batching thousands of requests, but underutilize hardware when run locally for a single user, whereas DiffusionGemma reverses that inefficiency by handing the processor a larger chunk of work at once.
The company was explicit that, because the model prioritizes speed and parallel generation, its overall output quality falls below standard Gemma 4, and it still recommends autoregressive Gemma 4 for work that demands maximum quality. DiffusionGemma's weights can be downloaded from Hugging Face, and the model runs with MLX, vLLM and Hugging Face Transformers. Working with NVIDIA, Google quantized the model for RTX 5090 and 4090 GPUs and optimized performance on Hopper and Blackwell systems with NVFP4 support. The release came as Google intensified its run of open-model launches in recent weeks, including Gemma 4 12B and Gemma 4 QAT.