{"id":430257,"date":"2026-06-11T09:00:09","date_gmt":"2026-06-11T03:30:09","guid":{"rendered":"https:\/\/dripp.zone\/news\/googles-diffusiongemma-ai-hits-1000-tokens-per-second-and-its-free-crypto-news-2\/"},"modified":"2026-06-11T09:22:44","modified_gmt":"2026-06-11T03:52:44","slug":"googles-diffusiongemma-ai-hits-1000-tokens-per-second-and-its-free-crypto-news-2","status":"publish","type":"post","link":"https:\/\/dripp.zone\/news\/googles-diffusiongemma-ai-hits-1000-tokens-per-second-and-its-free-crypto-news-2\/","title":{"rendered":"Google’s DiffusionGemma AI Hits 1,000 Tokens Per Second\u2014And It’s Free – Crypto News"},"content":{"rendered":"

<\/p>\n

\n
\n

In brief<\/h4>\n
    \n
  • Google released DiffusionGemma, a free open-weight model that generates entire 256-token blocks simultaneously via text diffusion\u2014hitting over 1,000 tokens per second on an NVIDIA H100, four times faster than standard autoregressive models.<\/li>\n
  • The custom drafter module DiffusionGemma needs for local inference doesn’t exist in any public runtime yet\u2014not in mlx-lm, not in LM Studio\u2014making it effectively unrunnable on most consumer setups today.<\/li>\n
  • On NVIDIA NIM, the model arrived preconfigured at 8,192 tokens of context\u2014below the 64,000-token floor that agentic frameworks like Hermes Agent require\u2014meaning autonomous workflows won’t run without manual reconfiguration.<\/li>\n<\/ul>\n<\/div>\n

    Google dropped DiffusionGemma today<\/a>, an open model AI that generates text the way image generators create pictures: start with noise, refine until it makes sense. It hits 1,000 tokens per second on an NVIDIA H100. (Tokens are the basic unit of information that an AI model handles.) That means it\u2019s four times faster than regular Gemma. It\u2019s also free, Apache 2.0, with weights on Hugging Face.<\/p>\n

    \n
    <\/figure>\n<\/div>\n

    The catch, as always, is in the fine print. Per Google’s announcement<\/a>, the model hits “700+ tokens per second on NVIDIA GeForce RTX 5090.” It also trails standard Gemma 4 on output quality.<\/p>\n

    Google says so themselves. This is a speed model, not a quality upgrade.<\/p>\n

    What this actually does<\/h2>\n

    Every LLM you’ve used is a typewriter. One token at a time with each word dependent on the last. That’s how autoregressive architectures work.<\/p>\n

    DiffusionGemma doesn’t do that. Instead of generating tokens sequentially, it starts with refined chunks of garbled text in parallel. Per Google’s developer guide<\/a>, it “starts with a canvas of random placeholder tokens” and iteratively locks in confident tokens until the whole block snaps into focus. Two hundred fifty-six tokens per forward pass. The GPU stays busy.<\/p>\n

    \n
    \"\"<\/figure>\n<\/div>\n

    The side effect is bidirectional attention\u2014every token can see every other token while being generated, which is impossible in autoregressive models (they cannot see the future, what is going to be encoded). That makes it unusually good at tasks where the end of the answer constrains the beginning: code infilling, structured output, constraint-heavy problems, etc. Google fine-tuned a version to solve Sudoku as a demo. The base model got roughly 0% of puzzles right.<\/p>\n

    The fine-tuned version hit 80%.<\/p>\n

    Text diffusion has been a research project for years. MDLM<\/a>, SEDD<\/a>, LLaDA<\/a>, Dream<\/a>\u2014academic models that proved the approach worked at small scales and mostly stayed as proof of concepts. Inception Labs shipped Mercury 2<\/a> in February 2026 as the first commercial diffusion reasoning model, claiming speeds five times faster than speed-optimized competitors.<\/p>\n