Google DeepMind's DiffusionGemma introduces a departure from traditional autoregressive language modeling by utilizing a non-autoregressive, block-diffusion approach to text generation. Unlike standard models that predict tokens sequentially from left to right, DiffusionGemma employs a denoising mechanism similar to image diffusion models. It begins with a canvas of noise tokens and iteratively refines entire blocks of 256 tokens in parallel until coherent text is formed. This architectural shift enables significant performance gains; on a single NVIDIA H100, the model achieves speeds exceeding 1,000 tokens per second, making it up to four times faster than standard Gemma models. Built upon a Gemma 26B Mixture-of-Experts (MoE) architecture, the model features 25 billion total parameters with fewer than 4 billion active parameters per token. It supports a 256K context window, accommodates over 140 languages, and processes multimodal inputs including text, image, and video. Released under the Apache 2.0 license, DiffusionGemma is optimized for tasks where the global structure of an answer is known upfront, though the author notes that its parallel nature may be less suited for tasks requiring strict, long-chain sequential reasoning compared to traditional autoregressive models.