langchain-apple-foundation-models

Use Apple's on-device Foundation Models as a drop-in LangChain chat model. Runs fully offline on macOS 26+ Apple Silicon -- no API key, no network call, no per-token cost.

from langchain_apple_foundation_models import ChatAppleFoundationModels

llm = ChatAppleFoundationModels()
llm.invoke("What is the capital of France?")

Install

pip install langchain-apple-foundation-models

Requires macOS 26+ with Apple Intelligence enabled and Apple Silicon.

Why

Apple shipped a Python SDK for its on-device model this year (apple-foundation-models), but there was no LangChain provider for it -- meaning none of LangChain's chains, agents, or tool-calling abstractions could target Apple's on-device model. This closes that gap.

Features

• Tool calling: pass LangChain @tool-decorated functions via .bind_tools([...]), same as any other LangChain chat model.
• Structured output: .with_structured_output(schema) for JSON-schema or Pydantic-constrained generation.
• Streaming: .stream(...) yields tokens as they're generated.
• Zero network calls -- generation happens entirely on-device via Apple's Neural Engine.

Architecture

flowchart LR
    App["Your LangChain code<br/>chains / agents / tools"] -->|".invoke() .stream() .bind_tools()"| Provider["ChatAppleFoundationModels"]
    Provider --> SDK["apple-foundation-models<br/>(Python SDK)"]
    SDK --> FM["Foundation Models framework<br/>on-device, Apple Silicon, Neural Engine"]
    FM -.->|"no network call, ever"| SDK

ChatAppleFoundationModels implements LangChain's standard chat model interface, so it drops into any existing chain, agent, or tool-calling setup as a swap-in for a cloud provider -- the only difference is generation happens on-device.

Example: tool calling

from langchain_core.tools import tool
from langchain_apple_foundation_models import ChatAppleFoundationModels

@tool
def get_weather(city: str) -> str:
    """Get the current weather for a city."""
    return f"The weather in {city} is sunny and 72F."

llm = ChatAppleFoundationModels().bind_tools([get_weather])
llm.invoke("What's the weather in Austin?")

Known limitations

• Multi-turn history is owned by the underlying Session, not LangChain's message list. Each ChatAppleFoundationModels instance lazily creates one on-device Session and reuses it across calls; if you pass a message list that diverges from what the session itself has tracked, the two can get out of sync. Fresh chains work naturally; manually-edited history does not yet.
• Built on the apple-foundation-models unofficial Python bindings (alpha), which itself wraps Apple's Foundation Models framework -- both inherit Apple's current 4,096-token context window.
• No support (yet) for Apple's newer multi-model / Private Cloud Compute routing announced at WWDC26 -- that requires an OS/SDK combination not yet stable enough to depend on. See the repo issues for status.

License

MIT