Google Releases DiffusionGemma 26B-A4B, an Open-Weight Text Diffusion Model

DiffusionGemma 26B-A4B-it uses discrete diffusion, with 25.2B total and 3.8B active parameters and a 256K-token context. It accepts text, images, and video and outputs text. Although it reaches about 1,000 tokens per second on an H100, up to four times autoregressive generation speeds, its MMLU Pro score of 77.6% and GPQA Diamond score of 73.2% are below comparable Gemma 4 models. The weights are on Hugging Face under Apache 2.0.

Visit original source