Google Introduces DiffusionGemma: Experimental Model Achieving 4x Faster Text Generation

Google DeepMind introduced DiffusionGemma, an experimental open model released under an Apache 2.0 license that fundamentally changes how text generation works. Unlike traditional autoregressive language models that generate text token-by-token from left to right, DiffusionGemma uses a diffusion-based approach similar to image generation models, producing entire blocks of up to 256 tokens simultaneously. The model achieves speeds of 700+ tokens per second on an NVIDIA GeForce RTX 5090 and over 1,000 tokens per second on an NVIDIA H100, representing approximately 4x faster inference than comparable autoregressive Gemma models. As a 26-billion-parameter Mixture of Experts model that activates only 3.8 billion parameters during inference, it fits within the 18GB VRAM limits of high-end consumer GPUs when quantized. The parallel generation approach enables bidirectional attention, allowing tokens to attend to all others simultaneously, which provides advantages for non-linear tasks like code infilling and mathematical problem-solving. Google acknowledges that DiffusionGemma prioritizes speed over output quality and recommends standard Gemma 4 models for applications requiring maximum quality.