Woofun AI reports that Google has integrated Multi-Token Prediction (MTP) into Pixel 9 and 10 series devices to accelerate onboard Gemini Nano v3. The architecture attaches a lightweight Transformer head to the main model, increasing inference speed by over 50% while maintaining security and output quality.

The zero-copy mechanism eliminates duplicate runtime memory overhead by allowing the prediction head to read the main model's cache via cross-attention, saving approximately 130MB of RAM. In scenarios like notification summaries, the system predicts nearly two extra tokens per inference, reducing processor wake-ups. Smart reply tasks show a 55% improvement in token acceptance rates.