A secure cloud database connects to AI tools, chat interfaces, and data services.
Azure HorizonDB, Microsoft's PostgreSQL-compatible database service for AI workloads, now has a feature that provisions AI models for you. The feature is called AI Model Management (AIMM). Microsoft lists it as a limited preview. That means you have to apply for access, so it isn't something you can switch on in any subscription today. If you have been wiring up model endpoints, keys and extensions by hand, AIMM is built to remove most of that work.

A secure cloud database connects to AI tools, chat interfaces, and data services. Access: limited preview, not open preview​

Microsoft's Learn documentation says AIMM is currently available as a limited preview for Azure HorizonDB, with access limited and requiring approval. To get in, you complete the AI Model Management limited preview participation form.

The wording on Azure's own status labels is easy to misread, so here is the distinction. Azure Updates defines "In preview" as available to all Azure customers for non-production use and testing. "In development" means a limited offering to select customers. The announcement headline carries the "In preview" tag, but the feature's own documentation describes a gated limited preview. Plan around the stricter reading. Treat AIMM as a pilot-only feature that needs approval.

Microsoft's announcement page gave no region list, enrollment deadline, SLA or production date, and I won't guess at any of them. Microsoft announced the wider HorizonDB service at Build. That announcement called it a public preview of an enterprise-ready, Postgres-compatible database service. It also invited readers to sign up to try AI model management in a limited preview.

What AIMM does​

Microsoft calls AIMM a premium feature in Azure HorizonDB that provides a fully managed experience for provisioning, configuring, and using AI models directly within your database. The documented behavior, once enabled, is:

  • It provisions a curated set of models for embedding, chat completion and reranking.
  • It installs and configures the azure_ai extension, with secure connections to the model endpoints.
  • It registers the models in the HorizonDB model registry under default aliases.
  • It folds billing and usage monitoring into the HorizonDB resource.

In practice this replaces a familiar chore list: deploy a model in Foundry, copy the endpoint, handle the key, then register it in the database. AIMM does that in one step.

The three default models​

AliasModelRole
default-embeddingtext-embedding-3-smallEmbeddings for semantic search and RAG
default-chatgpt-5.4Chat completion
default-rerankerCohere-rerank-v4.0-fastCross-encoder reranking of retrieved documents

Microsoft says HorizonDB automatically applies model updates, API version changes and lifecycle policies from Microsoft Foundry and reflects them in the registry. That is convenient, but it means you don't control when managed models change. Anyone who needs pinned model behavior for reproducibility should look at the bring-your-own-model route below. These model names come from today's documentation and may change.

How to enable it (after approval)​

The documented portal flow is short:

  1. Open your Azure HorizonDB instance in the Azure portal.
  2. In the resource menu, select AI Model Manager.
  3. Select the Enable managed models checkbox.
  4. Select Save.
  5. Review and accept the terms of use.

HorizonDB then creates the required Foundry resources, deploys the three managed models, installs azure_ai, registers the default aliases and sets up secure endpoint connections.

Using it from SQL​

Once AIMM is on, you can call AI functions with a default alias or leave the model parameter out. The functions then use the matching managed default. Microsoft's examples use ai.generate(), ai.embed(), ai.rank(), ai.extract() and ai.is_true().

Version boundary: those examples need azure_ai 2.5.6 or later. Microsoft says older versions and existing azure_ai.* or azure_openai.create_embeddings() callers should follow its migration guidance to the ai schema. If you copy older snippets from blog posts, some function names won't match. Check your extension version first.

Bring your own model and access control​

AIMM doesn't lock you into the managed set. You can register your own Microsoft Foundry deployments next to the managed ones. The registry stores each model's alias, name, type, endpoint URL, authentication type and endpoint key. You then select a model in SQL by its alias.

The documentation also covers restricting which database users can use specific models. It names a model_registry_manager role and the model_user_add, model_user_set and model_user_remove functions. This matters because registered models are open to all database users by default, so a security-minded team should review access after enabling AIMM.

Billing​

Microsoft states that AIMM pricing is based entirely on usage of the provisioned models. There is no extra charge for enabling the feature itself. Inference uses the same meters and rates as Microsoft Foundry, with no added markup. Costs show up on your Azure invoice and in Microsoft Cost Management, tagged to your HorizonDB resource.

The gain here is visibility, not savings. Embedding a large table with ai.embed() or running a chat model over many rows still costs money. Check current rates on the Foundry pricing page, and test on a small dataset first.

Where it fits: the semantic search tutorial​

Microsoft's tutorial for a semantic search app lists AIMM as a prerequisite. The workflow is:

  • Add vector and pg_diskann to the extension allow list, then run CREATE EXTENSION for each.
  • Generate embeddings with ai.embed() into a vector(1536) column.
  • Build a DiskANN index for approximate nearest-neighbor search.
  • Query by cosine distance, then rerank the top results with ai.rank().

The tutorial also warns that batch embedding can trigger throttling. Its advice is to tune the LIMIT value and retry after a minute if a statement fails partway through. If you skip AIMM, the documented fallback is to run CREATE EXTENSION azure_ai; yourself and register your own models.

What AIMM is not​

AIMM handles model provisioning and wiring. It does not guarantee good search relevance, chunk your documents or build a RAG application for you. HorizonDB's durable AI pipelines, which chain chunking, embedding, indexing, search and reranking, are a separate capability. Microsoft's pipeline documentation says that with AIMM enabled, models are provisioned and registered in the model registry automatically. The same documentation flags that model registration doesn't validate the endpoint or key during preview. That caveat applies to models you register yourself, so test your BYOM entries before relying on them.

Practical takeaways​

  • Request access early if you are evaluating HorizonDB for AI workloads, since approval is required.
  • Keep it non-production. Preview terms and the limited-preview status argue against production use.
  • Check your azure_ai version (2.5.6 or later) before using the ai.* examples.
  • Review model access in the registry, because all database users can use registered models by default.
  • Budget for inference. Usage is billed at Foundry rates through the HorizonDB resource.
  • Use BYOM if you need control over which model and version you run, given that managed models update automatically.

For teams already deep in PostgreSQL and Azure, AIMM looks like a sensible way to cut setup friction. The open questions are timing, regions and eventual production terms, and Microsoft hasn't answered those yet.

 

References

  1. (In preview) Limited Preview: AI model management in Azure HorizonDB Azure Updates 2026-11-17T23:42:10Z
  2. AI Model Management in Azure HorizonDB - Azure HorizonDB | Microsoft Learn learn.microsoft.com
  3. Build a Semantic Search Application in Azure HorizonDB - Azure HorizonDB | Microsoft Learn learn.microsoft.com