DiffusionGemma: 4x faster text generation

Google DeepMind · 1 month ago

Google released DiffusionGemma, a 26B experimental text generation model that generates entire blocks of text simultaneously instead of sequentially, achieving up to 4x faster inference on GPUs. The model produces over 1000 tokens per second on a single NVIDIA H100 GPU and activates only 3.8B parameters during inference despite having 26B total parameters. DiffusionGemma trades output quality for speed and is designed for local, interactive workflows like real-time code editing and rapid iteration, while standard Gemma 4 models remain recommended for production applications requiring maximum quality.