A drafter is a tiny, hyper-efficient model that runs alongside your “target” (or main) Gemma 4 model. By using a specialized speculative decoding architecture to decouple token generation from verification, these drafters deliver a 3x speedup without any degradation in output quality or reasoning logic.

By pairing the model with its drafter, developers are able to achieve:

— Improved responsiveness

— Supercharged local development

— Faster on-device performance

— Frontier-class reasoning without degradation

Video