Microsoft Foundry Local Hits GA: On-Device AI Ships to Production

Microsoft announced general availability of Foundry Local at Build 2026 on Tuesday, ending the preview period for the company's cross-platform runtime that runs full AI inference directly on a user's device. The release covers Windows, macOS, Android, and Linux, with automatic hardware routing across NVIDIA, AMD, Intel, Qualcomm, and Apple Silicon accelerators.

For developers who have spent the last two years sketching offline AI products on paper, the GA milestone is the green light to ship.

Why this matters

Most AI features today are network calls. The app sends your prompt to a cloud GPU, the model responds, the answer comes back. That model works — until it doesn't. Latency spikes when the network is weak. Privacy concerns block deployment in healthcare and legal contexts. Per-token costs make some product ideas economically impossible.

On-device AI fixes all three problems at once. The data never leaves the machine. Latency is whatever your local NPU can deliver. The marginal cost of each query is zero.

Foundry Local's pitch is that Microsoft has solved the hard part — making models actually run well across the messy reality of consumer and enterprise hardware — so application developers don't have to.

What was announced

Foundry Local is a runtime, a model catalog, and a packaging system. Developers integrate it as a library, choose a model from Microsoft's curated catalog, and ship. The runtime handles hardware detection at install time and routes inference to whichever accelerator is fastest on the user's machine — GPU, NPU, or CPU.

Microsoft says supported hardware includes NVIDIA GPUs from the 2000-series forward, AMD GPUs from the 6000-series forward, AMD NPUs, Intel iGPUs and NPUs, Qualcomm Snapdragon X Elite chips, Qualcomm NPUs, and Apple Silicon across Mac and iPad. The runtime falls back to CPU if no accelerator is detected.

Packaging is meant to be invisible to end users. Microsoft says Foundry Local is "small enough to bundle directly inside application installers, with zero dependencies." That means an ISV can ship a single executable that brings its own AI runtime — no separate model download step, no Python environment to install, no GPU driver to configure.

Microsoft also previewed WSL 3 alongside Foundry Local, with near-native GPU and NPU passthrough for Linux workloads running on Windows. Together, the two enable the same model code to run identically on a developer's laptop, a Linux container, and a user's machine.

Industry impact

The audience that matters most here is ISVs — independent software vendors building AI features into shrink-wrapped products.

For healthcare software vendors, Foundry Local makes ambient documentation and decision support tools shippable without sending patient data to a cloud. For legal tech vendors, contract review and drafting can happen entirely on a lawyer's laptop. For developer tool makers, local-first coding assistants become practical for environments where SaaS isn't allowed — defense contractors, intelligence agencies, regulated financial institutions.

The broader business model implication is that AI features can finally be sold the way software has always been sold: with a license fee, not a per-token meter. That changes both economics and product design.

It also intensifies competitive pressure on cloud AI providers. Every workload that runs locally is a workload that doesn't generate API revenue. Microsoft is betting that the volume of new offline AI applications will more than offset the cloud workloads that move on-device, and that capturing the developer relationship at this layer is strategically more important than maximizing inference revenue in the short term.

Expert perspectives

Microsoft engineers at Build emphasized the practical engineering work behind GA. The hardest problem in shipping cross-platform local AI isn't the model — it's the long tail of edge cases across thousands of GPU and NPU combinations, driver versions, and operating system updates. Foundry Local's "zero detection code required" promise is Microsoft's claim that it has absorbed that complexity.

Independent developers reacting on social media flagged the WSL 3 integration as the underrated detail. Local model development workflows that previously required dual-booting or VM tricks now work natively, which lowers the barrier for the Linux-first developer community that has historically resisted Windows AI tooling.

Analysts noted the strategic parallel with Apple's CoreML and Google's MediaPipe. Microsoft is positioning Foundry Local as the de facto cross-platform runtime, betting that developers will prefer one runtime that works everywhere over three platform-specific stacks.

What's next

Foundry Local is downloadable today through the Foundry Local GitHub repository and the Microsoft Foundry portal. The model catalog launches with options across chat, audio, and vision, including distilled versions of MAI models tuned for on-device performance.

Microsoft has signaled that the on-device variants of MAI-Voice-2 and MAI-Image-2.5 will be added to the Foundry Local catalog over the summer. Expect partner ISVs to start shipping Foundry Local-powered features in fall 2026 product releases.

Pricing for the runtime itself is free. Models in the catalog are licensed under their original terms, with most carrying commercial use rights.

Bottom line

Foundry Local turns on-device AI from a research project into a shipping product. If you build software that runs on customer machines and you've been waiting for the right moment to add real local AI, that moment is now. The economics, the privacy posture, and the latency profile all favor going local where you can. Microsoft just removed the engineering excuse for not trying.