The Catalogue's Two Families: Sold by Azure, and From Partners and Community
The Catalogue's Two Families: Sold by Azure, and From Partners and Community
What this slice covers
Foundry Models is the catalogue you browse when you need a model, and it is large — well over ten thousand entries, growing by roughly fifty a month, spanning foundation models, reasoning models, small language models, multimodal models, domain-specific models, and industry models. Faced with that, the first structural fact to learn is not any individual model's name. It is that the catalogue is divided into exactly two commercial families, and that division determines who supports you, who bills you, and what guarantees you get.
The two families are Foundry Models sold by Azure, and Foundry Models from partners and community.
Models sold by Azure
Microsoft hosts these models and sells them under Microsoft Product Terms. They come from a range of providers — Azure OpenAI models among them, alongside selected models from other top providers — but the commercial relationship is with Microsoft throughout.
What that buys you is a consistent bundle: Microsoft provides the support; the models are deeply integrated with Azure services and infrastructure; they pass internal review against Microsoft's Responsible AI standards; documentation and transparency reports describe their risks, mitigations, and limitations; and they carry enterprise-grade scalability, reliability, security, and service level agreements. Some also offer fungible provisioned throughput, meaning reserved capacity you can move across models in the family rather than pinning to one.
Billing runs through Azure meters as first-party consumption services.
Models from partners and community
Here is the fact that reorders most people's intuition: this second family is not a small annexe to the first. These models make up the vast majority of the Foundry Models. They come from trusted third-party organisations, partners, research labs, and community contributors — Anthropic's Claude family and the open models from the Hugging Face hub are the named examples.
The characteristics differ point by point from the first family. They are developed and supported by those external parties rather than by Microsoft. They cover a far wider spread of specialised and niche capabilities. Validation is typically done by the providers themselves, with Azure supplying integration guidance. Innovation arrives quickly, because community and partner release cycles are not gated on Microsoft's. And support and maintenance are the provider's responsibility, with SLA levels that vary.
Billing for this family runs through Azure Marketplace under commercial marketplace terms rather than as first-party Azure consumption — a small detail that carries real procurement consequences.
Choosing between the families
The documentation frames the decision along three axes, and they are worth learning as a set:
- Use case and requirements. Deep Azure integration, guaranteed support, and enterprise SLAs point to models sold by Azure. Specialised or innovation-led scenarios point to partners and community.
- Support expectations. One family gives you a single Microsoft support path; the other gives you the provider's, with variable structures.
- Innovation and specialisation. Rapid access to niche capability from leading labs and emerging providers is the partner family's distinctive strength.
Note what is not on that list: raw benchmark quality. Both families contain state-of-the-art models. The split is commercial and operational, not a quality ranking, and treating it as a ranking is the classic misreading.
Responsibilities that apply either way
Whichever family a model belongs to, the obligations on you are identical: comply with the law in how you use it; read the model description in the catalogue, the provider's model card, and the related documentation; select a model appropriate to your use case; and put appropriate safeguards in place so your use meets Microsoft's acceptable use policy and code of conduct for enterprise AI services. Choosing a Microsoft-supported model does not transfer any of that to Microsoft.
If a model you want is absent, the catalogue itself has a request path: search for the name in the portal's model catalogue, and when nothing matches, the option to request a model appears.
Mistakes to avoid
- Assuming everything in the catalogue is Microsoft-supported. Most of it is not.
- Assuming models sold by Azure are all Microsoft-built. They come from multiple providers; Microsoft's role is hosting, integration, and commercial ownership.
- Reading the split as a quality tier list rather than a support-and-billing distinction.
- Expecting a uniform SLA across the partner family. Levels vary by provider.
- Forgetting that responsible-use obligations sit with you in both families.
What to carry forward
Two families, one dividing line: who hosts, supports, and bills. Sold by Azure means Microsoft on all three counts, with SLAs to match. Partners and community means the provider, with breadth and speed as the trade. Everything else in this topic — deployment options, quota rules, retirement policy — is applied through that distinction.