grabbou/qwen3-0.6b-CoreAI

Qwen3 0.6B for Core AI (iOS)

Qwen/Qwen3-0.6B exported to Apple Core AI for iOS 27 with the apple/coreai-models recipe, mixed 4-bit and 8-bit palettization, static 4096-token context. It runs through Foundation Models' LanguageModelSession with the CoreAILM product of coreai-models.

This is a copy of the ios/ export in mlboydaisuke/qwen3-0.6b-CoreAI-official at 7406ece, unchanged. Core AI exports depend on the OS and toolchain that produced them, so this pins the artifact that @react-native-ai/core-ai is tested with.

Files

ios/ holds metadata.json, the exported model (qwen3_0_6b_mixed_4bit_8bit_static.aimodel, 440 MB) and the tokenizer.

Use

With @react-native-ai/core-ai, download the ios/ folder and pass its path:

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

const model = coreAI.languageModel({
  id: 'qwen3-0.6b',
  source: { type: 'file', uri: qwen3ModelPath },
  variant: 'iOS',
})

const { text } = await generateText({
  model,
  prompt: 'Explain on-device AI in one sentence.',
  reasoning: 'none',
})

Qwen3 thinks before answering unless reasoning is off. With thinking on, set maxOutputTokens: generating past the 4096-token context crashes Apple's iOS engine.

Export

The original export command, from apple/coreai-models:

uv run coreai.llm.export qwen3-0.6b --platform iOS

License

Apache-2.0, as the original model. Qwen3 is by the Qwen team, Qwen3 Technical Report.