Google released a fast text model that runs on your own machine
Google released a fast text model that runs on your own machine. DiffusionGemma — an open-weights model (Apache 2.0, free on Hugging Face) — generates text up to 4x faster than other Gemma models and runs locally on a single GPU.
What it actually means
The interesting part isn't the speed number — it's where it runs. DiffusionGemma is open-weights and free to download, so for the right job you stop paying an API per token at all. It's experimental and built for speed over precision, so it won't replace your frontier model on hard reasoning. But for the high-volume, low-stakes work — autocomplete, drafting, inline edits — running something local could cut a real line off your bill. Worth a test on one noisy, repetitive task this week. It pairs directly with the cost squeeze everyone's feeling: the cheapest token is the one you don't rent.
An open model you host yourself is a token bill you stop paying.
4xfaster text generation than Google's other Gemma models
Source: Google · Monday 29 June 2026