Google DeepMind Releases DiffusionGemma 26B-A4B Under Apache 2.0
Google DeepMind released DiffusionGemma 26B-A4B on Hugging Face, Kaggle, and Vertex AI Model Garden. The 26B MoE activates about 4B parameters and uses text diffusion to generate over 1,000 tokens per second on one H100, up to four times faster than autoregressive models. Google positioned it for speed-sensitive experiments and recommended Gemma 4 for production.
Visit original source