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

Google has introduced DiffusionGemma, an experimental open-source text generation model that applies diffusion techniques — previously used in image generation — to large-scale language modeling. Rather than predicting tokens sequentially left-to-right, the model starts with a 256-token block of random placeholders and iteratively refines the entire block in parallel, locking in high-confidence positions and re-evaluating uncertain ones across multiple passes. Benchmark data shows the FP8 version reaching approximately 1,008 tokens per second on a single H100 and 1,288 on an H200, representing roughly five to six times the speed of a standard autoregressive baseline under optimal single-user conditions, and approximately 4x faster than comparable Gemma 4 models in Google's own benchmarks. As a 26B Mixture of Experts model that activates only 3.8B parameters during inference, it fits within 18GB VRAM when quantized, making it compatible with consumer GPUs such as the NVIDIA RTX 4090 and 5090. The bidirectional attention architecture gives the model a structural advantage for constrained generation tasks — such as code infilling, structured data generation, and problems requiring context from later in a sequence — as demonstrated by a fine-tuned version achieving an 80% success rate on Sudoku puzzles that the base model could not solve at all. However, Google and independent researchers caution that the speed advantage diminishes in high-throughput cloud serving where autoregressive models already saturate compute through batching, and that overall output quality remains below standard Gemma 4 for open-ended generation tasks. The model is available now on Hugging Face and is the first diffusion language model natively supported in the open-source vLLM inference platform.