Text of code.visualstudio.com › docs › intelligentapps › models as Styx crawled it on Nov 18, 2026. The page may have changed since. Only its words are kept, without pictures or layout, and only the first part of a long page.

Open the live page

Explore and manage AI models in Foundry Toolkit

AI agents: See llms.txt for documentation discovery and navigation. Documentation articles are also available as Markdown by appending .md to the article URL. Join VS Code at Microsoft Ignite or catch the sessions online. A year of VS Code, wrapped. Explore the highlights. Explore and manage AI models in Foundry Toolkit View as Markdown Ask in VS Code Ask in VS Code Insiders Foundry Toolkit helps you prepare a model for application development in Visual Studio Code.

Start by deciding where inference should run, and then deploy a cloud model, download a supported local model, or connect to an existing endpoint. This article explains how to find and set up a model, test it, and understand who owns the deployment, connection, or downloaded data. Supported model sources The publisher of a model is not necessarily its hosting provider. For example, a model published by one organization might be hosted in Microsoft Foundry, by the publisher, or in a local runtime.

Choose a source based on where inference must run and which resource you want to manage. For example, use a Foundry deployment for a shared cloud endpoint, connect a provider-hosted API that your application already uses, or select a local runtime to test a model on supported device hardware. Model source or runtime Platform Setup action Your responsibility Microsoft Foundry Remote inference on supported extension platforms Deploy a catalog model to a selected Foundry project and resource.

Manage the project, deployment, quota, access, and service configuration. OpenAI, Anthropic, or Google Provider-hosted inference on supported extension platforms Connect a supported model with provider credentials. Maintain the provider account, credentials, and any provider-side resources. NVIDIA NIM Provider-hosted inference on supported extension platforms, with supported local NIM models on Windows Connect to a hosted endpoint or download a supported local model.

Maintain the endpoint or local runtime and its model data. Custom or bring your own model (BYOM) Any compatible endpoint reachable from the extension host, including localhost and private networks Connect with a full Chat Completions URL, model details, and a credential. Operate the endpoint and verify which capabilities it implements. Ollama Windows, macOS, and Linux Connect to an installed Ollama runtime and add a model.

Maintain Ollama, its endpoint, and downloaded models. Foundry Local SDK Windows x64 and Arm64, and macOS Arm64 Download and cache a supported model variant. Maintain the local cache and meet the model's hardware requirements. ONNX and ONNX Converted Windows x64 and Arm64 Download a supported model or add output from the model conversion tool. Maintain the local model data and choose a model variant supported by your hardware.

Windows AI APIs, including Phi Silica Supported Windows devices Use an operating-system-backed model shown under Microsoft Foundry on Windows. Meet the Windows and hardware requirements. Availability depends on the device. On Windows, Microsoft Foundry on Windows groups supported Foundry Local SDK models and entries labeled For Windows ML or For Windows AI APIs . On macOS, the Toolkit displays the supported SDK integration as Foundry Local .

The broader Foundry Local SDK supports other environments, but this table describes the Foundry Toolkit integration. Find a model The model catalog combines models that you can deploy, download, or connect. Use its filters to narrow the catalog without treating catalog metadata as a guarantee that a model is available for your account, project, hardware, or endpoint. Distinguish image input from image generation when you choose a model.

A chat model can analyze images without producing image output. To generate images in Toolkit, select a supported Foundry image-generation model and prepare its deployment. Select the Foundry Toolkit view in the Activity Bar. Go to My Resources > Models . Select the Catalog tab. Search by model name or description, or use the catalog filters: Hosted by identifies the service or runtime that serves the model. Publisher identifies the organization that publishes the model.

Capability filters identify advertised features, such as attachments or structured output. Confirm model and endpoint support when you test the model. Local model filters reflect available runtimes and model variants. CPU, GPU, or NPU eligibility depends on the model, device, and installed runtime. Open a model card and review its source, available actions, prerequisites, and license information. Searches and narrow filters can change which catalog sections are visible.

Clear the search and filters to return to the broader catalog. The following example filters the catalog to the Phi-4 model family and shows the availability and hosting-source filters. An administrator can also connect models for a configured Foundry project. These administrator-connected models are separate from catalog-created deployments and connections that you add yourself. They are not another Catalog provider or a deployment that you manage from this page.

Their availability depends on the project and model-selection workflow. Use Agent Builder to select a model while building an agent, and contact your administrator if an expected connection is unavailable. Deploy a model to Microsoft Foundry Deploy a catalog model when you want Microsoft Foundry to host inference in your Foundry project. Before you start, make sure that: You are signed in and can access the target Foundry project and resource.

The model is available in the selected region. The model supports the selected deployment type, and the required quota and capacity are available. You accept any publisher agreement that the deployment flow displays. For a deployment that uses compatible settings for the selected project, open the model card's Deploy menu and select Deploy with default settings when that action is available. This action starts deployment with the Toolkit's defaults.

To review and customize the deployment settings: In My Resources > Models > Catalog , open the model that you want to deploy. Open Deploy and select Custom deploy . Confirm the read-only Project and Connected AI resource fields. Select the intended project in Foundry Toolkit before you open this form. Confirm the model and set the deployment name, deployment type, model version, version upgrade policy, tokens-per-minute quota, and content filter available for that model.

Review the settings and resolve any inline validation errors. Review any publisher agreement that appears before you continue. Select Deploy to Microsoft Foundry to submit the deployment. After the resource list refreshes, go to My Resources > Models > Foundry to find the deployment. The following example shows the custom deployment form before submission. Project and resource names are redacted. The result is a cloud inference deployment in the selected Foundry resource.

It is not an agent resource. Open the deployment in the Playground to test inference, or select it in Agent Builder when you build an agent. Region, quota, pricing, authentication, and safety configuration remain service-specific. Review the Foundry model deployment prerequisites and options , Foundry model quota and limits , and Foundry cost planning guidance before you create a production deployment. The Foundry deployment overview explains service-side deployment options and their content filtering and networking boundaries.

Add or connect a model Use the setup flow that matches who hosts the model. A connection in Foundry Toolkit makes an existing endpoint available to development tools. It does not create or operate the provider's deployment. Connect a provider-hosted model In My Resources > Models > Catalog , open a supported model from OpenAI, Anthropic, Google, or NVIDIA. Select the action to add or connect the model. Enter the provider credential and any endpoint or model details requested by the setup prompts.

Compl

The rest of this page is on the live site.