DeepMind's DiffusionGemma is 4x faster
Google DeepMind's DiffusionGemma, out June 10, generates text by diffusion and claims up to 4x faster output, with NVIDIA tuning it for local PCs.
On June 10, Google DeepMind released DiffusionGemma, an experimental open-weights addition to the Gemma family that generates text by diffusion, producing whole blocks of text at once instead of the usual one-token-at-a-time decoding. Google’s claim is up to 4x faster text generation on GPUs than autoregressive models, with reported throughput above 1,000 tokens per second on a single H100 and 700 or more on a GeForce RTX 5090. The weights are on Hugging Face under Apache 2.0: a 26 billion parameter mixture-of-experts model with 3.8 billion active at inference.
The release came with immediate ecosystem support: NVIDIA published optimizations the same day for running DiffusionGemma locally on RTX hardware, pitching it for on-device assistants and drafting tools.
Why this matters
Speed is the boring constraint that decides whether local AI is usable. A 4x jump changes what a laptop-grade machine can do without a cloud subscription: drafting, summarizing, and rewriting at speeds that feel instant instead of stuttering. For operators who keep sensitive text off cloud services, that matters more than benchmark points.
The caveats are right in Google’s own post, which is refreshing: overall output quality is lower than standard Gemma 4, the speedup is aimed at local and low-concurrency use rather than busy cloud serving, and memory-bandwidth-bound machines like Apple Silicon Macs see less benefit. Test it on your real workload before moving anything. If you want the mental model for why this architecture differs, what an LLM actually is covers the standard decoding loop this replaces, and the solo operator AI stack covers where a fast local model actually fits.