Google Releases DiffusionGemma, a 26B MoE Model Offering 4x Faster Text Generation

Google has launched DiffusionGemma, an experimental open-source language model released under an Apache 2.0 license that uses text diffusion to generate entire blocks of text simultaneously rather than token by token. The 26B Mixture of Experts model activates only 3.8B parameters during inference, enabling it to run within 18GB VRAM on high-end consumer GPUs when quantized. It achieves over 1,000 tokens per second on an NVIDIA H100 and 700+ on an RTX 5090, making it up to four times faster than standard autoregressive models. The model supports bidirectional attention across 256 parallel tokens, offering advantages for tasks like code infilling, inline editing, and non-linear text structures. Google notes that DiffusionGemma is intended for researchers and speed-focused workflows, and that standard Gemma 4 remains the recommended choice for production-quality output.