Foundry Local

Foundry Local is a lightweight runtime and set of native SDKs for building applications that run AI entirely on a user's device, so conversations never leave the device, responses start without network latency, and the app keeps working offline with no per-token cost or API keys. The runtime handles model acquisition, hardware acceleration, model lifecycle management, and inference through ONNX Runtime, automatically selecting the best available NPU, GPU, or CPU on the end user's hardware.

It ships a curated catalog of models optimized and quantized for on-device use, covering chat completions from families such as GPT OSS, Qwen, DeepSeek, Mistral, and Phi, and audio transcription with Whisper, with models versioned so an application can pin a version or update automatically. It exposes an OpenAI-compatible API, including the Responses API format, so an application already built against the OpenAI SDK can point at a Foundry Local endpoint with minimal changes, and an optional local server can serve models to multiple processes or tools such as LangChain.