[john-rocky]

Summary

Adds @react-native-ai/litert, an AI SDK provider for LiteRT-LM, built on react-native-litert-lm. It runs .litertlm models on iOS and Android, with generation, streaming, tool calling and structured output.

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

const model = litert('hf://litert-community/gemma-4-E2B-it-litert-lm')
await model.prepare()

const { text } = await generateText({ model, prompt: 'Write a haiku about coding.' })

Sampling and tools apply when the model loads, so changing them between calls reloads the model and returns a warning. Options the runtime doesn't support return an unsupported warning.

Given a Hugging Face repo, the provider picks the file: the GPU-only file on iOS and the full bundle on Android, because the GPU-only Gemma 4 E2B file gave wrong text, with no error, on a Pixel 11 Pro.

Docs are in website/src/content/docs/docs/litert/, and the root README gets a row and a section.

Test plan

• Unit tests, typecheck and lint pass.
• In an Expo app on the packed package, from a clean install with litert-community/gemma-4-E2B-it-litert-lm, 30 checks pass on a Pixel 11 Pro (Android 17) and on an iPhone 18 Pro (iOS 27.2). They cover generation, streaming, tools, structured output, conversations, and these failures: no such file or repo, a broken or truncated file, an abort during a load and during a call, a tool call that doesn't parse (injected), tool arguments that fail the schema, and a prompt that's too long.
• Not verified on a device: backend: 'npu', speculativeDecoding, a repo with a litertlm_manifest.json, a gated model's accessToken, and Android asking for memory back.
• The provider isn't in apps/example or harness yet.

[vercel]

@john-rocky is attempting to deploy a commit to the Callstack Team on Vercel.

A member of the Team first needs to authorize it.