Google Research demonstrated Gemini Nano acceleration on Pixel via frozen Multi-Token Prediction (MTP). The trick: predict multiple tokens at once without extra training, keeping the base model unchanged. Coverage was mostly the official technical writeup — no consumer buzz — with community threads (a GPT2-in-BASIC port) as background color. Small in noise, meaningful in trajectory: the on-device LLM cost curve keeps bending. Portability to non-Pixel hardware and measured power savings are the next signal points.
Gemini × Infrastructure & Hardware
Google speeds Gemini Nano on Pixel
Google speeds Gemini Nano on Pixel
✎ Story body
▲ Official & Press
Official
Accelerating Gemini Nano models on Pixel with frozen Multi-Token Prediction
Google speeds up Gemini Nano on Pixel via frozen multi-token prediction
Community
GPT2-BASIC: Portable Machine Intelligence in BASIC
GPT2-BASIC: a portable GPT-2 implementation written in BASIC