Updated

Updated · heise online · Jul 16

Google Releases Gemma-4 Diffusion Model Hitting 5,000 Tokens a Second as Quality Lags

Updated

Updated · heise online · Jul 16

Google Releases Gemma-4 Diffusion Model Hitting 5,000 Tokens a Second as Quality Lags

1 articles · Updated · heise online · Jul 16

Google’s new Gemma-4 discrete-diffusion model generated 4,000 to 5,000 tokens per second in tests on an Nvidia RTX 5090, far above roughly 300 tokens per second for the standard model.
The speed comes from speculative decoding via discrete diffusion, which can produce up to 256 tokens at once and is aimed mainly at draft prediction rather than final responses.
Response quality was weak in the reported tests: the model often produced tokens without displaying a usable answer, limiting its appeal outside acceleration workflows.
Google’s release adds a rare text-focused diffusion model to the open Gemma-4 family, while rival approaches such as DeepSeek’s DSpark claim even bigger gains but still lack mature software support.

Sources

New AI models are achieving incredible speeds, but can our strained power grids and supply chains actually sustain this AI revolution?

As AI costs approach human developer salaries, is re-hiring experienced engineers becoming the smarter business decision?

With US export controls now limiting AI access, is local, open-source AI the only path to digital independence for other nations?

1 total