Foundry Local Review, Pricing & Alternatives (2026)

Foundry Local is Microsoft’s end-to-end solution for embedding local AI features in applications that run on a user’s device. Microsoft announced its general availability on April 9, 2026, making 2026 its documented launch year. The product combines a runtime, SDKs, a curated model catalog, and optional local service components, with support for Windows, Linux, macOS, and other device types described by Microsoft. Its SDK is available for C#, JavaScript, Python, and Rust, while the repository also documents C and C++ interfaces. Microsoft describes the runtime as approximately 20 MB, separate from downloaded model files. The public repository records 2,596 stars, 377 forks, and 128 open issues, and as a Microsoft product it has no separate funding or valuation.

A typical integration starts by initializing the SDK, inspecting the available execution providers and catalog entries, then requesting a model by identifier or alias. The runtime downloads the selected package, reports progress, loads it into device memory, and exposes chat or transcription calls. Chat responses can stream incrementally, while applications can combine speech to text, tool calls, and responses through the same SDK. When the application exits, it can unload the model; later launches reuse the local cache rather than repeating acquisition. The SDK runs in the application process, so end users do not need the CLI. For multi-process access or existing integrations, developers can instead configure the optional server and send OpenAI-format requests, including chat completions, audio transcriptions, streaming, and tool definitions.

Its underlying stack is a native Foundry Local Core implemented in C++, exposed through a stable C ABI and language bindings, with ONNX Runtime handling inference. The SDK can register execution providers such as CUDA, QNN, OpenVINO, TensorRT, and WebGPU, while Windows uses Windows ML and macOS uses Metal through WebGPU. The October 23, 2026 version 2.2.1 release updated ONNX Runtime to 1.31.0 and ONNX Runtime GenAI to 0.18.0, added catalog version 2, coding-model support, Linux ARM64 CUDA registration, stricter tool-call validation, and cleanup for damaged caches. CUDA and cuDNN must be installed separately. These choices make deployment dependent on provider libraries and platform configuration, while versioned catalog metadata helps match models to compatible runtime versions.

Foundry Local has no subscription, Azure requirement, per-token fee, or documented hosted usage tier for device execution. The free SDK is MIT licensed, but that does not grant equivalent rights to every catalog model. Model-specific licenses may govern redistribution, commercial use, attribution, or other obligations, so shipping a model requires reviewing its catalog metadata and license rather than treating the runtime license as sufficient. The roughly 20 MB runtime also understates distribution needs: models download separately, remain cached on the user’s storage, and may require additional provider components. Hardware is therefore an indirect cost. Microsoft does not publish a full list of supported GPUs, NPUs, or CPUs, and some CUDA deployments require separately installed CUDA and cuDNN runtimes. Teams should budget storage, bandwidth, device eligibility, licensing review, and packaging work instead of assuming software is the only expense.

Foundry Local’s curated catalog limits arbitrary checkpoint use, so teams needing broad model selection or frequent custom-model testing may prefer Ollama or Lemonade. Ollama provides a simpler local experimentation workflow, while Lemonade offers a community-oriented server and AMD-focused control. Windows ML is a lower-level alternative for Windows applications, giving developers more direct platform integration but requiring more responsibility for runtime details. Foundry Local’s SDK can avoid installing its CLI for end users, yet its optional CLI and REST interface remain preview components, with possible breaking changes. First-run model downloads, cache storage, hardware compatibility, and individual model licenses still need to be handled in deployment planning. Choose Foundry Local for supported-language applications that need embedded, offline inference across Windows, macOS, and Linux with managed model acquisition.