Google has released DiffusionGemma, an experimental open-source 26B mixture-of-experts model that generates text using diffusion techniques instead of traditional left-to-right token-by-token processing. By drafting entire 256-token blocks simultaneously, it achieves up to 4x faster inference compared to autoregressive models. The model activates only 3.8B parameters during inference and fits within 18GB VRAM, making it compatible with consumer GPUs like the Nvidia RTX 5090. It features bidirectional attention and self-correction via confidence scoring, making it well-suited for code infilling, inline editing, and non-linear text tasks. Released under Apache 2.0, it's available on Hugging Face, GitHub, vLLM, and Google Cloud Model Garden. Key trade-offs include lower output quality than standard Gemma 4 and diminishing returns in high-QPS cloud serving environments.
