Features: Azure Local Foundry

Scope

This is the features page for Azure Local Foundry, one of the three targets in ADR-0011. Compare all three targets on the Deployment targets hub.

This target has not been built yet

Nothing on this target has been deployed from this repository, and no automation exists for it yet. Everything below is drawn from accepted decisions and first-party research. Treat it as a design, not an as-built record.

What is decided today

• Concurrent multi-user serving is the design point, unlike the single-host target.
• Agentic Retrieval is available, and it requires a GPU and the gpt-oss-20b class.
• Disconnected operation is supported via the Azure Local disconnected operations appliance, version 2604.3.0 or later.
• Entra ID authentication is unavailable on the Helm onboarding channel, which is why that channel is disqualified.
• No content filter or responsible-AI guardrail is documented at all. This is the governance-relevant gap: a workload moved here from Azure AI Foundry silently loses every control ADR-0007 relies on.
• Vision input is available here and nowhere else on premises:pixtral-12b-2409 and three nemotron-nano-12b-v2-vl variants, vLLM and GPU only (SPIKE-22).
• No Foundry Agent Service, no MCP surface, no fine-tuning service, and no batch job API. A custom model can be brought from an OCI registry via spec.model.custom. The one asynchronous primitive is the predictive server's queue, which returns queue_depth and retry_after when full.
• No embeddings and no text to speech. Structured outputs (json_schema) are not documented (SPIKE-31).

The research and decisions behind this page

These were open when this page was first written. They are now written and linked below. Where a finding contradicted an earlier claim on this page, the page has been corrected and says so inline.

• SPIKE-31, the cross-track feature parity spike