Google Speeds Pixel 9 and 10 Gemini Nano by 50% With 130MB-Lighter MTP
Updated
Updated · Google Research · Jun 26
Google Speeds Pixel 9 and 10 Gemini Nano by 50% With 130MB-Lighter MTP
1 articles · Updated · Google Research · Jun 26
Summary
Pixel 9 and 10 devices have already received Google’s retrofit of Multi-Token Prediction onto frozen Gemini Nano v3 models, lifting on-device AI generation speeds by 50% or more.
The upgrade targets the one-token-at-a-time bottleneck in mobile LLMs by adding a lightweight MTP head to the existing model instead of using a separate drafter, preserving output and safety behavior bit-for-bit.
Google said the design cross-attends to the main model’s KV cache rather than building its own, cutting runtime memory use by up to 130MB per instance and avoiding extra prompt-processing latency.
In production features such as AI Notification Summaries and Proofread, the system predicts nearly two extra tokens per inference pass on average, reducing verification steps and energy use.
The rollout extends Google’s push to make private, on-device AI more practical on phones, with future work aimed at parallel decoding, branching token paths and looser verification for further edge-device gains.
With Google boosting Pixel's AI speed by 50%, how will Apple's next iPhone respond in the on-device AI race?
Google's MTP solves the speed problem. What's the next barrier for powerful AI agents to run entirely on phones?
As on-device AI rivals the cloud, is Google making its own high-margin cloud services obsolete?
Pixel 10 Series and Gemini Nano: How Google’s On-Device AI Sets a New Benchmark for Smartphone Performance, Privacy, and Longevity in 2026
Overview
Google's Pixel 10 series introduces the powerful Tensor G5 chip, marking a major leap for the Pixel ecosystem. This custom silicon is designed to boost performance and user experience, especially by enabling advanced on-device AI. Gemini Nano, Google's on-device large language model, runs directly on the Pixel 10 without needing cloud access, making AI features faster and more private. Together, the Tensor G5 and Gemini Nano allow the Pixel 10 to deliver smarter, more responsive features, highlighting Google's focus on custom hardware and AI to set new standards for smartphones.