Google DiffusionGemma: 1,000 Tokens/Sec on Nvidia GPUs

Google DeepMind has released DiffusionGemma, a new text-generation model that creates text in parallel blocks. This model reaches up to 1,000 tokens per second on Nvidia GPU hardware. Benchmarks show DiffusionGemma runs four times faster than previous Gemma models on equivalent computing power. Another report confirmed ten times higher token throughput in long-context inference tests on Nvidia hardware. Standard large language models generate one token at a time. DiffusionGemma generates entire text blocks simultaneously, using a diffusion-based architecture. This approach sharply reduces latency for long outputs. The model also self-corrects complex markdown and structured formats during generation. This capability is aimed at developers building code assistants, documentation tools, and structured data pipelines. It's optimized for local deployment on Nvidia RTX consumer GPUs and DGX enterprise systems. This development could mean faster and more efficient text generation for many applications.