Skip to main content
Arka can route LLM requests to Apple’s on-device Foundation Models (Apple Intelligence) through the apple-fm provider. Inference stays on your Mac — no API key required.

Requirements

  • macOS 26+ (Apple Intelligence enabled in System Settings)
  • Apple Silicon Mac that supports Apple Intelligence
  • Python optional extra: pip install 'arka-agent[apple-fm]' (installs apple-fm-sdk)
On Linux and Windows, the provider is registered but always skipped gracefully.

Quick setup

Set Apple Intelligence as your preferred provider:
Or add to .env:

How it works

  1. Native SDK — Arka calls apple-fm-sdk directly (SystemLanguageModel + LanguageModelSession).
  2. Fallback chain — On supported Macs, apple-fm is added after Ollama in the auto-built fallback chain.
  3. Streaming — Supported via the SDK’s stream_response() when available.
Check availability anytime:

Optional: apple-fm-cli server

If the SDK is not installed but you run an OpenAI-compatible local server (for example apple-fm-cli), Arka auto-detects it:
Configure Arka to use that endpoint (default port 8765 avoids conflict with vLLM on 8000):
The server does not need a real API key; Arka sends a placeholder bearer token.

Troubleshooting

Disable the provider without uninstalling:
  • Model hosting setup — Ollama, vLLM, LM Studio, Exo
  • arka llm providers --models — list configured providers
  • arka hybrid status — combine local and hosted routes