Accelerating Gemini Nano Models on Pixel With Frozen Multi-Token Prediction
Google, Friday, June 26th, 2026
Google Research retrofits Multi-Token Prediction onto frozen Gemini Nano for ~50% faster on-device generation.
Google Research announced a method to retrofit Multi-Token Prediction onto frozen Gemini Nano v3 models, enabling faster on-device text generation on Pixel 9 and 10 phones.
The approach attaches a lightweight transformer head to the main model rather than using a separate drafter, eliminating memory redundancy through a zero-copy architecture that shares the backbone's cached states.
It achieves roughly 50% speedups while reducing energy consumption. Real-world applications like AI Notification Summaries and Proofread generate text faster without fine-tuning separate models per task.