🚀 Foundry Local v2.5.0 - Bring AI closer to your code
Local coding takes center stage in Foundry Local SDK v2.5.0. Build coding experiences around models that run on the developer's device, with support announced for four models across the Qwen, MAI Code, and Nemotron families.
💻 Meet the local coding lineup
Choose a model for your application and build around local inference with Foundry Local's SDKs. This release's coding lineup brings together:
| Model | Catalog name |
|---|---|
| Qwen3.8 27B | qwen3.8-27b
|
| MAI Code 1.1 Flash | mai-code-1.1-flash
|
| Nemotron 3.5 Bolt 74B-A9B | nemotron-3.5-bolt-74b-a9b
|
🛠️ Start building with the SDK
Use the SDK to discover compatible model variants, manage downloads, and run inference inside your application.
Start with the SDK quickstarts, then explore the SDK samples to build your local coding experience. Model visibility and usable variants depend on hardware and catalog availability.
🔥 Catch us at Microsoft Ignite
Join the team at Microsoft Ignite to explore what's possible with on-device AI:
- Breakout - BRK255: Build private, responsive on-device AI agents with Foundry, with Sam Kemp and Baiju Meswani.
- Lightning Talk - LTG1267: AI at Your Fingertips: Foundry Local on Windows, with Wayne Chuang.
✨ Also in 2.5.0
- More faithful model behavior. Native tool continuations preserve model-formatted tool history on supported protocols. Model-configured sampling defaults are honored unless the caller overrides them.
- Bring your own model catalog. Discover and download compatible models from an Azure ML registry, then run them locally using the existing model lifecycle. See the registry catalog guide.
- GPU-capacity-aware model discovery. The catalog filters out GPU variants when the configured execution route has known total VRAM below the model's declared requirement. Unknown capacity remains visible; this check does not guarantee that a model will load.
🐛 Bug Fixes
- Platform reliability fixes. Unicode model paths work correctly on Windows, macOS shared-host WebGPU shader compilation is fixed, and WebGPU provider teardown is ordered correctly during manager disposal.
- Catalog task alignment. The SDK now uses the catalog metadata task names
decisionandrankingfor decision and ranking models. - Multimodal WebGPU compatibility. Fixed a provider-selection conflict that could prevent multimodal WebGPU models from running when the CUDA ExecutionProvider was registered.
📚 Keep exploring
Documentation | Report an issue | 2.5.0 release-branch changes
💙 A huge thank you to our contributors
This release is a team effort! A huge thank you to @aciddelgado, @apsonawane, @baijumeswani, @bmehta001, @EmmaNingMS, @GopalakrishnanN, @hanbitmyths, @jiafatom, @kunal-vaishnavi, @prathikr, @rui-ren, @sayanshaw24, @selenayang888, and @tianleiwu for helping make local coding possible.
And to everyone testing builds, reporting issues, and sharing feedback: thank you for helping bring local coding to life!
Want to help build what's next? Bug reports, documentation improvements, and pull requests are welcome. Read the contribution guide and help shape the next release.