Details
- NVIDIA congratulated Google DeepMind on the launch of DiffusionGemma and highlighted the model’s parallel generation approach.
- The company said DiffusionGemma generates 256 tokens in parallel per step.
- NVIDIA reported performance of 150+ tokens per second on DGX Spark and 1,000+ tokens per second on a single H100.
- Support is available from day one through BF16 and NVFP4 checkpoints on Hugging Face.
- NVIDIA also pointed to an NVFP4 version on Hugging Face, indicating immediate ecosystem support for deployment and experimentation.
- The post frames the launch as a joint milestone for model efficiency and accelerated inference across NVIDIA hardware.
Impact
The announcement signals that diffusion-based language models are moving from research novelty toward hardware-supported deployment. By pairing Google DeepMind’s launch with BF16 and NVFP4 checkpoints on Hugging Face, NVIDIA is positioning its stack to capture early developer attention and make the model easier to test on mainstream inference hardware. The performance figures on DGX Spark and H100 also underscore how specialized GPU optimization is becoming a competitive differentiator as model providers look to reduce latency and increase throughput.
