Generating

Generate text with Gemini Nano: pass android() to the AI SDK’s generateText or streamText.

Text Generation

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

const { text } = await generateText({
  model: android(),
  system: 'You are a helpful travel assistant.',
  prompt: 'Suggest three things to do in Kraków.',
})

System messages go to ML Kit as its system instruction.

Options

temperature, topK, seed and maxOutputTokens go to ML Kit. topP, stopSequences, presencePenalty and frequencyPenalty return an unsupported warning.

Streaming

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

const result = streamText({
  model: android(),
  prompt: 'Write a short poem about the sea.',
})

for await (const delta of result.textStream) {
  console.log(delta)
}

Pass an abortSignal to stop a generation. The native request is cancelled too.

Structured Output

Not supported: Output.object and responseFormat: { type: 'json' } return an unsupported warning, and the model answers in text. ML Kit only generates structured output for Kotlin classes compiled with its KSP processor, not for a JSON Schema at runtime.

Conversations

Pass the conversation as messages. The provider sends the whole history to Gemini Nano on every call, as user and model turns, so it keeps no session of its own. Keep the history within the context window.

Note

Gemini Nano versions without system prompt support (the systemPrompt feature), such as nano-v2 on a Pixel 9 Pro, read only the first text part of a request. On those, the provider merges the system instruction and the text of every turn into one prompt, with the images first and each turn labeled by role, so the model sees the whole conversation. Models with system prompt support get the instruction and the turns separately.

const { text } = await generateText({
  model: android(),
  messages: [
    { role: 'user', content: 'My name is Ada.' },
    { role: 'assistant', content: 'Nice to meet you, Ada!' },
    { role: 'user', content: 'What is my name?' },
  ],
})

Image Prompts

Gemini Nano takes up to 10 images per request. Pass bytes, base64, or a local file:// or content:// URL:

const { text } = await generateText({
  model: android(),
  messages: [
    {
      role: 'user',
      content: [
        {
          type: 'file',
          mediaType: 'image/jpeg',
          data: new URL('file:///path/to/photo.jpg'),
        },
        { type: 'text', text: 'What is in this photo?' },
      ],
    },
  ],
})

Other file types, such as audio, are rejected with INVALID_MESSAGE.

Reasoning

Where the device’s Gemini Nano supports thinking (features includes 'thinking'), any reasoning level other than 'none' turns it on, and the thoughts come back as reasoning parts. Gemini Nano’s thinking is on or off, with no budget. Without thinking support, the request runs without it and returns an unsupported warning for reasoning.

const result = streamText({
  model: android(),
  reasoning: 'high',
  prompt: 'Is 17 a prime number?',
})

providerOptions.android.enableThinking turns it on or off when reasoning is not set.

Context Window

The context size depends on the device’s Gemini Nano, so read it at runtime:

const { contextSize } = await android.getModelInfo()

To keep long chats inside it, wrap the model with wrapLanguageModelWithHistory, which keeps a rolling window and summarizes older messages. See Conversation History.

ML Kit reports no token usage, so usage is empty. Count a prompt’s tokens with countTokens:

const tokens = await android().countTokens('How long is this prompt?')