Do I have to download a model?

Not for Apple and Gemini Nano: they use models that ship with the operating system. The system downloads some assets, such as Apple’s embedding and transcription models, on first use. Core AI and Llama run models you ship or download.

Which devices does it run on?

Apple text generation needs iOS 26 on a device with Apple Intelligence turned on; embeddings and speech work on older versions. Gemini Nano needs a physical Android device with AICore support. Llama runs on iOS and Android.

Does anything leave the device?

No, unless you opt in. Inference runs locally. Apple’s Private Cloud Compute model and ADK’s cloud Gemini are cloud models you choose explicitly.

Which AI SDK version do I need?

React Native AI 0.13 and above works with AI SDK v7. Use 0.12 for v6 and 0.11 or below for v5. Every provider needs polyfills for streams and structuredClone.

Can I test in the iOS Simulator?

Yes for Apple: the Simulator uses your Mac’s Apple Intelligence once it is turned on in macOS. Core AI and Gemini Nano need a physical device.

Can I mix on-device and cloud models?

Yes. They are all AI SDK models, so you can pick one per call, and fall back to a cloud provider when apple.isAvailable() is false.

On-device AI for your mobile app. Every platform’s models, one API.

Run the models each platform provides, or your own, right on your users’ devices. Drop-in Vercel AI SDK providers for React Native: swap the model per platform and keep the rest of your code. No server, no API key.

Get started

$npm install @react-native-ai/apple ai

React Native AIzsh

$ npm install @react-native-ai/apple ai

# the Apple provider and the AI SDK

React Native AITripPlanner.tsx

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

const result = streamText({
  model: apple(),
  prompt: 'Plan a 3-day trip to Kyoto',
})

for await (const delta of result.textStream) {
  setReply((text) => text + delta)
}

9:41 Airplane mode

Plan a 3-day trip to Kyoto

Day 1: Fushimi Inari at sunrise, then Gion in the evening.

Day 2: Arashiyama bamboo grove and the Tenryū-ji gardens.

Day 3: Nishiki Market, then Kiyomizu-dera for the view.

apple() · on device

[01] How it works

Swap the model. Keep the code.

01
Install the providers
Add the provider for each platform you ship to, next to the AI SDK.
$ npm install @react-native-ai/apple ai
02
Pick a model per platform
Use apple() on iOS and adk() on Android, or a model you ship with Core AI or Llama.
$ npm install @react-native-ai/adk
03
Call the AI SDK
Pass the model to generateText, streamText, embed or any other AI SDK function. The rest of your code stays the same, on iOS and Android.
$ npx expo run:android

React Native AIapple.ts

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

const result = streamText({
  model: apple(),
  prompt: 'Summarize my notes',
})

@react-native-ai/appleFoundation Models, provided by the platform: native and fast.Set up Apple

React Native AIcore-ai.ts

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

const result = streamText({
  model: coreAI.languageModel({
    id: 'qwen3-0.6b',
    source: qwen3ModelDirectory,
  }),
  prompt: 'Summarize my notes',
})

@react-native-ai/core-aiCore AI, provided by the platform: native and fast.Set up Core AI

React Native AIadk.ts

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

const result = streamText({
  model: adk(),
  prompt: 'Summarize my notes',
})

@react-native-ai/adkGemini Nano via AICore, provided by the platform: native and fast.Set up ADK

React Native AIllama.ts

import { downloadModel, llama } from '@react-native-ai/llama'
import { streamText } from 'ai'

const path = await downloadModel(
  'bartowski/Qwen_Qwen3.5-2B-GGUF/Qwen_Qwen3.5-2B-Q4_K_M.gguf'
)
const model = llama.languageModel(path)
await model.prepare()

const result = streamText({ model, prompt: 'Summarize my notes' })

@react-native-ai/llamallama.cpp via llama.rn, through the third-party llama.rn binding.Set up Llama

[02] Runtimes

The fastest runtime on each platform. One API over all of them.

Apple, Core AI and ADK run on the AI stack the platform provides: native, hardware-accelerated and with its system models. Llama runs GGUF models sized for phones on iOS and Android through llama.rn, a third-party binding.

Native · by Callstack

Apple

Runtime: Foundation Models
Runs on: iOS 26+
Models: Built in: Apple Intelligence

The model that ships with iOS, plus embeddings, transcription and speech from Apple’s own frameworks.

@react-native-ai/appleNative · by Callstack

Core AI

Runtime: Core AI
Runs on: iOS 27+
Models: Yours: exported or hosted .aimodel bundles

Apple’s runtime for models you ship: text, image prompts, image generation, transcription, embeddings and vision tasks.

@react-native-ai/core-aiNative · by Callstack

ADK

Runtime: Gemini Nano via AICore
Runs on: Android 14+
Models: Built in: Gemini Nano

Google’s on-device Gemini through the Agent Development Kit, with tool calling and an opt-in cloud Gemini.

@react-native-ai/adkThird-party binding

Llama

Runtime: llama.cpp via llama.rn
Runs on: iOS and Android
Models: GGUF models from Hugging Face that fit in memory

The portable option: an AI SDK provider over the third-party llama.rn binding, for GGUF models small enough for a phone, on both platforms.

@react-native-ai/llama

[03] Small context windows

Long chats, small models.

On-device models have a few thousand tokens of context. Wrap any model to keep a rolling window, summarize older turns and drop finished tool calls, without changing your chat state.

Conversation history

React Native AIchat.ts

import { apple } from '@react-native-ai/apple'
import { wrapLanguageModelWithHistory } from '@react-native-ai/utils'
import { streamText } from 'ai'

const model = wrapLanguageModelWithHistory(apple(), {
  rollingWindowMessages: 12,
  dropCompletedToolCalls: true,
})

const result = streamText({ model, messages })

[04] What you get

On the device. Off the network.

Data stays local

Inference runs on the device, so prompts and replies never reach a server.

Works offline

No network round trip and nothing to wait for. It works in airplane mode.

No keys, no bills

System models need no API key, and there is no per-token cost.

Cloud is opt-in

Private Cloud Compute and cloud Gemini run only when you ask for them.

[06] FAQ

Do I have to download a model?: Not for Apple and Gemini Nano: they use models that ship with the operating system. The system downloads some assets, such as Apple’s embedding and transcription models, on first use. Core AI and Llama run models you ship or download.
Which devices does it run on?: Apple text generation needs iOS 26 on a device with Apple Intelligence turned on; embeddings and speech work on older versions. Gemini Nano needs a physical Android device with AICore support. Llama runs on iOS and Android.
Does anything leave the device?: No, unless you opt in. Inference runs locally. Apple’s Private Cloud Compute model and ADK’s cloud Gemini are cloud models you choose explicitly.
Which AI SDK version do I need?: React Native AI 0.13 and above works with AI SDK v7. Use 0.12 for v6 and 0.11 or below for v5. Every provider needs polyfills for streams and structuredClone.
Can I test in the iOS Simulator?: Yes for Apple: the Simulator uses your Mac’s Apple Intelligence once it is turned on in macOS. Core AI and Gemini Nano need a physical device.
Can I mix on-device and cloud models?: Yes. They are all AI SDK models, so you can pick one per call, and fall back to a cloud provider when apple.isAvailable() is false.

Get started

On-device AI for your mobile app. Every platform’s models, one API.

Swap the model. Keep the code.

Install the providers

Pick a model per platform

Call the AI SDK

The fastest runtime on each platform. One API over all of them.

Apple

Core AI

ADK

Llama

Long chats, small models.

The AI SDK you know, on the device.

Text and streaming

Structured output

Tool calling

Embeddings and reranking

Transcription and speech

Image generation

Generative UI

DevTools

On the device. Off the network.

Data stays local

Works offline

No keys, no bills

Cloud is opt-in

On-device AI for your mobile app. Every platform’s models, one API.

Swap the model. Keep the code.

Install the providers

Pick a model per platform

Call the AI SDK

The fastest runtime on each platform. One API over all of them.

Apple

Core AI

ADK

Llama

Long chats, small models.

The AI SDK you know, on the device.

Text and streaming

Structured output

Tool calling

Embeddings and reranking

Transcription and speech

Image generation

Generative UI

DevTools

On the device. Off the network.

Data stays local

Works offline

No keys, no bills

Cloud is opt-in

Frequently asked questions