Models & Availability

Gemini Nano comes with AICore, Android’s system service for on-device models. AICore downloads it on first use and updates it on its own.

Checking Availability

getModelInfo() checks with AICore:

import { android } from '@react-native-ai/android'

const info = await android.getModelInfo()
// {
//   availability: 'available',
//   baseModelName: 'nano-v2',
//   contextSize: 8192,
//   features: ['systemPrompt', 'thinking', 'structuredOutput', 'functionCalling'],
// }

availability	Meaning
available	Ready to use
downloadable	Supported, and downloads with prepare()
downloading	Downloading now
unavailable	Not supported on this device
unknown	AICore didn’t answer

baseModelName, contextSize and features are only set once the model is available.

Preparing

prepare() downloads Gemini Nano when needed and loads it into memory, so the first request answers sooner. Generating prepares on its own when you don’t, and logs a warning.

await android().prepare({
  onProgress: ({ phase, bytesDownloaded, bytesToDownload }) => {
    console.log(phase, bytesDownloaded, bytesToDownload)
  },
})

Choosing a Model

ML Kit’s ModelConfig picks between Gemini Nano releases:

import { createAndroidProvider } from '@react-native-ai/android'

const preview = createAndroidProvider({
  model: { releaseStage: 'preview', preference: 'fast' },
})

• releaseStage: stable (default) or preview, the newest Gemini Nano for devices in the AICore developer preview
• preference: full (default) favors quality, fast favors latency

Tool calling needs Gemini Nano 4: the preview release with fast, or the default model on devices that ship it, such as the Galaxy Z Flip 8 and Fold 8. getModelInfo takes the same selection.