Gemma 4 inference gets faster through multi-token prediction drafters that generate several tokens at once

The article says Gemma 4 is being accelerated with multi-token prediction drafters, a technique intended to speed up inference by proposing multiple tokens per step. Details in the provided body are limited, but the focus is clearly on improving model serving efficiency.

Faster inference is increasingly central as AI providers try to reduce latency and serving costs without sacrificing output quality. Techniques like drafting point to a broader trend of optimizing model deployment, not just scaling model size.