Why distribute the text?
While the AI research community has explored broadcast-based text generation for many years, applying it to large-scale models remains a challenge. DiffusionGemma changes this by changing the way models use hardware.
Trade in traditional models
Most language models act like a typewriter, producing one token at a time from left to right. In the cloud, this works well because servers can aggregate thousands of user requests to share the hardware load. But when used locally for a single user, this word-for-word process leaves your dedicated GPU or TPU underutilized – it spends most of its time waiting for the “keyboard.”
DiffusionGemma reverses this inefficiency. Instead of predicting the words one by one, it writes a complete category of 256 tokens at a time. By giving the computer processor a large portion of the work at once, DiffusionGemma uses your hardware to its full potential. It improves the predictability of your model from a single, step-by-step typewriter to a large press that stamps an entire document at once.
