====== Google Nano (Gemini Nano) ======

Gemini Nano is Google smallest and most efficient AI model, designed to run natively on Android devices without requiring cloud connectivity. Unlike Gemini Pro or Flash which run on Google cloud infrastructure, Nano executes AI inference directly on the user device, delivering zero network latency, no API costs, and strong privacy guarantees because user data never leaves the device. ((source [[https://developer.android.com/ai/gemini-nano|Android Developers - Gemini Nano]]))

===== Technical Architecture =====

Gemini Nano consists of two variants optimized for different device capabilities:

^ Variant ^ Parameters ^ Target ^
| Nano-1 | 1.8 billion | Low-memory devices |
| Nano-2 | 3.25 billion | Higher-memory devices |

Both variants use 4-bit quantization and are created through distillation from larger Gemini models, allowing them to inherit advanced capabilities while fitting mobile hardware constraints. ((source [[https://localaimaster.com/blog/gemini-nano-android-guide|Local AI Master - Gemini Nano Guide]]))

The model runs within Android AICore system service, which leverages device hardware to enable low inference latency and automatically keeps the model updated. ((source [[https://developer.android.com/ai/gemini-nano|Android Developers - Gemini Nano]]))

===== Performance =====

On Pixel 10 Pro devices, the latest Gemini Nano v3 achieves:
  * **940 tokens per second** for text-to-text processing
  * **0.6 seconds** for image-to-text encoding
  * **87 percent** of basic AI tasks run locally without cloud calls
  * **40 percent less battery** usage compared to cloud-based alternatives

((source [[https://techpurs.com/gemini-nano-on-android-2026-what-it-does-simple/|TechPurs - Gemini Nano on Android 2026]]))

===== On-Device Capabilities =====

**Text Processing:**
  * Summarization of documents up to 3,000 words into bullet points (English, Japanese, Korean)
  * Smart replies with context-aware suggestions in Google Messages, WhatsApp, Line, and KakaoTalk via Gboard
  * Text rewriting with tone adjustment (formal, casual, excited)
  * Proofreading and text correction

**Multimodal Capabilities (Pixel 9 and later):**
  * Image description and alt-text generation for accessibility through TalkBack
  * On-device speech recognition powering Pixel Recorder and Call Notes
  * Real-time audio analysis for features like Scam Detection

**Pixel-Exclusive Features:**
  * Pixel Screenshots: AI-powered screenshot search
  * Call Notes: Automatic transcription and summarization
  * Scam Detection: Real-time fraud detection during calls
  * Pixel Recorder Summaries: 3-bullet summaries of recordings over 30 minutes

((source [[https://localaimaster.com/blog/gemini-nano-android-guide|Local AI Master - Gemini Nano Guide]]))

===== Supported Devices =====

^ Device Category ^ Models ^ Nano Version ^
| Google Pixel | Pixel 8, 8 Pro, 8a | Nano v1 (text-only) |
| Google Pixel | Pixel 9, 9 Pro, 9 Pro XL, 9a | Nano v2 |
| Google Pixel | Pixel 10 series | Nano v3 (full multimodal, Tensor G5) |
| Samsung | S24, S25 series, Galaxy Z Fold/Flip 6 | Variable versions |
| Other Android | Xiaomi, Motorola, compatible processors | Variable versions |

Devices require sophisticated chips such as Google Tensor G4 or later, Qualcomm Snapdragon 8 Gen 3, or equivalent processors to support Gemini Nano. ((source [[https://blogs.infosys.com/digital-experience/emerging-technologies/emergence-of-on-device-ai-introducing-gemini-nano.html|Infosys - Gemini Nano]]))

===== Privacy Benefits =====

  * **Full on-device processing:** No internet connection required
  * **Complete offline operation:** AI features function without any network connectivity
  * **Data never leaves the device:** Inherent privacy safeguards for sensitive information
  * **No per-request charges:** Free for users with supported devices
  * **Blocks third-party data scraping:** By design, no external data access

These characteristics make Gemini Nano ideal for processing sensitive data like personal messages, health information, and confidential documents. ((source [[https://gemilab.net/en/articles/gemini-dev/gemini-nano-on-device-ai-guide|GemiLab - Gemini Nano Guide]]))

===== Developer Integration =====

Developers can integrate Gemini Nano through **ML Kit GenAI APIs**, which provide high-level interfaces for summarization, proofreading, rewriting, image description, speech recognition, and a low-level Prompt API for flexibility. ((source [[https://developer.android.com/ai/gemini-nano|Android Developers - Gemini Nano]]))

The **Android AI Edge SDK** provides access to Gemini Nano APIs without additional subscription costs. Third-party app support is expanding, though platform-native apps like Gboard, Recorder, and TalkBack already leverage Nano capabilities. ((source [[https://www.abhs.in/blog/google-gemini-nano-android-developers-what-you-can-build-2026|Abhishek Gautam - Gemini Nano Developers 2026]]))

===== Latest Developments (2025-2026) =====

  * **Pixel 10 Release:** Gemini Nano v3 launched with substantially improved performance and full multimodal support
  * **Enhanced ML Kit GenAI APIs:** New APIs for easier implementation of summarization, proofreading, rewriting, and image description
  * **Performance gains:** Text-to-text processing reaching 940 tokens per second on Pixel 10 Pro
  * **Chrome integration:** CPU support for Gemini Nano rolling out in Chrome 140, expanding on-device AI beyond Android ((source [[https://developer.chrome.com/blog/gemini-nano-cpu-support|Chrome Developers - Gemini Nano CPU Support]]))
  * **Android 2026 integration:** Gemini Nano ships baked into Android system-level AICore with 2.8x faster response times versus 2025 hybrid models ((source [[https://techpurs.com/gemini-nano-on-android-2026-what-it-does-simple/|TechPurs - Gemini Nano on Android 2026]]))

===== Gemini Nano vs Cloud Models =====

^ Aspect ^ Gemini Nano ^ Gemini Pro/Flash (Cloud) ^
| Processing location | On-device | Google cloud servers |
| Internet required | No | Yes |
| Latency | Near-instant | Network round-trip |
| Cost per request | Free | Per-API-call pricing |
| Privacy | Data stays on device | Data sent to servers |
| Capabilities | Focused (text, images, speech) | Full range (reasoning, generation, tools) |
| Model size | 1.8B - 3.25B parameters | Much larger |

===== See Also =====

  * [[gemini_fast_thinking_pro|Gemini Flash, Thinking, and Pro]]
  * [[chatgpt_claude_gemini_comparison|ChatGPT, Claude, and Gemini Comparison]]
  * [[google_ai_video_models|Google AI Video Models]]
  * [[google_veo|Google VEO 3]]

===== References =====