Gemini Nano on Android

Gemini Nano on Android

by Google

Google’s on-device model for Android, delivered through the AICore system service for private, offline inference.

Facts (Android developer docs, 2026-10-01)

• Gemini Nano runs in AICore, which manages model weights and updates, LoRA adapters, safety filters, hardware acceleration and request isolation. The OS distributes and updates the model, so apps ship no weights.
• Main access route: ML Kit GenAI APIs: Prompt (text-only or multimodal), Summarization, Proofreading, Rewriting, Image Description (all Beta per developers.google.com/ml-kit/genai, rechecked 2026-10-02) and Speech Recognition (Alpha; basic mode API 31+, advanced Gemini Nano mode only Pixel 10-11). A android skills add ml-kit-genai-prompt-api CLI route is documented.
• Privacy: AICore follows Private Compute Core principles (restricted package binding, internet only via Private Compute Services, no storage of inputs/outputs after processing).
• Cost: no per-call server cost; usage is quota-limited per app; ML Kit GenAI additional terms apply.

What changed in 2026

• Gemini Nano 4 (nano-v4): “built on the architecture foundation of the recently released Gemma 4 model” (gemma-4; Android Developers Blog, July 2026). The ML Kit Prompt API needs a specific Nano version: nano-v2 (OnePlus, OPPO, POCO and others), nano-v3 (Pixel 9-10, Galaxy S26 and other recent flagships), nano-v4 (Pixel 11 series and Galaxy Z Fold8 series) (ML Kit GenAI docs, 2026-10-02). The task-specific APIs (Summarization, Proofreading, Rewriting, Image Description) list many devices across Pixel 9-11, Galaxy S25-S26, OnePlus, OPPO, Xiaomi and others.
• Scale: the Android blog says Gemini Nano 4 runs on over 140 million devices.
• Prompt API is Beta in the ML Kit docs (earlier secondary reports dated the alpha-to-beta move to 2026-01-28; not rechecked).
• Firebase AI Logic hybrid inference (on-device with cloud fallback; modes PREFER_ON_DEVICE, ONLY_ON_DEVICE, PREFER_IN_CLOUD, ONLY_IN_CLOUD): per the Firebase docs (2026-10-02) on-device inference is supported only for web apps on Chrome desktop (Chrome Prompt API); it is not available on Android through Firebase AI Logic. The earlier “Android SDK specifics” question is answered: not supported.
• No embedding API appears in the ML Kit GenAI docs (the earlier “embedding API coming soon” claim was removed).
• Not carried over because only a secondary source (9to5Google, 2026-08-27) supported them: “Gemini Intelligence” device requirements (12 GB RAM) and the Galaxy Z Flip8 / Fold8 Ultra device list.

See also

Open on-device alternative: gemma-4.

Open items

• Nano 4 model size and benchmark figures are not published in the pages checked; per-region availability not verified. These are not key claims.

Sources (accessed 2026-10-01 and 2026-10-02)

• https://developer.android.com/ai/gemini-nano
• https://developers.google.com/ml-kit/genai
• https://android-developers.googleblog.com/2026/07/android-on-device-inference.html
• https://9to5google.com/2026/08/27/gemini-intelligence-nano-4/
• https://firebase.google.com/docs/ai-logic/hybrid-on-device-inference
• Search summary: Prompt API alpha to beta 2026-01-28 (developer.android.com blog / ML Kit docs snippets)