Gemini Nano

Gemini Nano is Google’s compact Gemini model built to run directly on certain Android devices. Instead of sending every request to a cloud data center, the phone can handle selected language and multimodal tasks locally—useful when responsiveness, privacy, or connectivity matters.

What runs on the phone
Gemini Nano is an on-device large language model: a neural network whose learned weights are stored on the device and executed by its local CPU, GPU, or NPU (neural processing unit). It is much smaller than Google’s cloud-hosted Gemini models, because a handset has finite RAM, battery capacity, and cooling. The trade-off is deliberate: Nano handles bounded, practical tasks well, rather than trying to match a large cloud model on every kind of reasoning or writing.

How Android uses it
On supported devices, Android exposes Gemini Nano through system components such as AICore and developer-facing APIs. A feature can pass a prompt—and, for models that support it, local content such as text or images—to the model, then receive generated text or structured results. Typical uses include:

• Summarizing a recorded conversation or long text without uploading it.
• Drafting context-aware reply suggestions in messaging tools.
• Describing or organizing content for accessibility and productivity features.
• Powering assistant features when the device has no usable network connection.

Why its edge design matters
Keeping inference on the handset can reduce round-trip delay and keeps sensitive inputs under the device’s local security boundary. But it does not make computation free: generation consumes energy, uses memory for both model weights and the growing KV cache, and can be slowed by thermal throttling. Developers therefore design short prompts, limit output length, and provide fallbacks for unsupported phones. Gemini Nano is a practical example of a modern language model shaped around a phone’s real physical limits, not just its benchmark score.