Skip to content

Available models: the two on-premises targets

This page was split in two

It used to cover both on-premises targets in one place, and that made the two rosters harder to tell apart rather than easier. They are two catalogs, not one. Each now has its own page:

The third target has always had its own page: Available models: Azure AI Foundry.

Why they had to be separated

The rosters diverge in both directions, which a single merged table consistently understated:

Foundry LocalAzure Local Foundry
Shared ONNX core35 aliases35 aliases
vLLM roster (100 entries)noyes
NPU and alternate-accelerator variantsyesno
Vision inputnoyes
Azure subscription requirednoyes
Shapea ~20 MB library inside your appan Arc-enabled Kubernetes extension

Only the middle is common: 35 aliases out of 170 total entries. A reader who wants to know what one target can run should not have to filter out the other target's rows to find out.

This corrected the premise of ADR-0017 decision 5, which assumed the two differed only in execution provider and deployment mechanics.

See also