Google Releases Open Diffusion Language Model DiffusionGemma
DiffusionGemma marks a practical step toward faster, open-weight language models that can handle tasks requiring repeated rethinking, which autoregressive models struggle with.
Key Facts
- Google released DiffusionGemma on June 10, 2026, an open diffusion language model that generates text roughly four times faster than the Gemma 4 26B A4B model at a similar scale.
- DiffusionGemma has 25.2 billion total parameters with 3.8 billion active under a mixture-of-experts architecture.
- A fine-tuned variant of DiffusionGemma demonstrated high accuracy on Sudoku puzzles, a task that requires repeatedly revisiting the same cell.
- DiffusionGemma is released under the Apache License 2.0 and is available for download on Hugging Face.
- NVIDIA collaborated on a quantized NVFP4 version of DiffusionGemma that runs on a single GeForce RTX 5090.
Reporting from 1 source: GIGAZINE.
Google has released DiffusionGemma, an open diffusion language model that generates text about four times faster than its Gemma 4 series while maintaining competitive accuracy. The model uses a diffusion process rather than the standard autoregressive approach, and a fine-tuned version shows strong performance on iterative reasoning tasks like Sudoku.
Google released DiffusionGemma on June 10, 2026, an open diffusion language model that generates text roughly four times faster than the company's own Gemma 4 26B A4B model at a similar scale. Instead of the standard autoregressive approach that outputs tokens one by one, DiffusionGemma iterates over the entire sequence of noise to converge on the final text, a method borrowed from image generation AI.
The model has 25.2 billion total parameters with 3.8 billion active under a mixture-of-experts architecture. Google says it builds on the earlier Gemini Diffusion research announced in May 2025. A fine-tuned variant of DiffusionGemma demonstrated high accuracy on Sudoku puzzles, a task that requires repeatedly revisiting the same cell, which the company highlighted as a strength of the diffusion approach.
DiffusionGemma is released under the Apache License 2.0 and is available for download on Hugging Face. NVIDIA collaborated on a quantized NVFP4 version that runs on a single GeForce RTX 5090.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.