Diffusion Gemma

DiffusionGemma is an experimental generative AI model from Google DeepMind that uses discrete diffusion rather than conventional token-by-token autoregressive generation. Built on the Gemma 4 architecture, it generates and repeatedly refines blocks of text in parallel. It is designed for speed-critical applications such as interactive AI, coding, editing, code infilling and other real-time workflows. The model has 26B total MoE parameters but activates about 3.8B during inference, making quantized local deployment possible on consumer GPUs. It is an open-weights model released under Apache 2.0. Google currently provides access through Hugging Face, Kaggle and Vertex AI. No country-specific unavailability is stated in the official documentation reviewed.