Source: Microsoft Foundry — Foundry Models Overview

Foundry's model catalogue: models sold by Azure vs partner/community models, serverless vs managed-compute deployment, billing and content-safety differences.

Contents

Summary

Explains the two catalogue categories and the two deployment options.

Key claims

  • 10,000+ models, about 50 added per month, in two categories: models sold by Azure (Microsoft-hosted, Microsoft Product Terms, Microsoft support and SLAs, some with fungible PTU) and models from partners and community (e.g. Anthropic Claude, Hugging Face).
  • Claude in Foundry is governed by Anthropic’s data processing addendum and commercial terms.
  • Serverless deployment: Microsoft-hosted API billed per token, with partner models billed through Azure Marketplace. Managed compute: weights on dedicated VMs billed per core-hour, which needs a hub-based (classic) project.
  • Serverless language models get default Content Safety filters (billed separately); embedding and time-series models have none. Filtering applies only through the Azure OpenAI v1 APIs unless you add Content Safety yourself.
  • Serverless endpoints follow the resource’s public-network-access flag; hubs with private endpoints created before July 11, 2024 need re-creation.

Feeds

Azure AI Foundry

Suggest changes →