Foundry Models

The catalog changes monthly. Resolve exact versions, SKUs, registration and regional availability live through microsoft-foundry and the Azure-direct models (opens in a new tab) and region availability (opens in a new tab) pages; never copy a version from memory.

Starting points by task

Recommend the first fit, then offer alternatives only when cost, region, quota or registration blocks it. Prefer GA for production and flag Preview or registration-required models before deploying.

Joint region and capacity preflight

Treat placement as one subscription-specific decision, not a static model lookup. Before persisting AZURE_LOCATION or declaring a deployment, validate every required resource in each candidate region:

Start with the preferred region, then use the target subscription’s live availability and quota checks through the maintained microsoft-foundry guidance. If any required resource fails, reject that candidate and rank alternatives that satisfy the complete resource set; do not move one resource independently unless the architecture explicitly supports cross-region placement.

Write the placement decision to ./.azure/deployment-plan.md:

When the deployed location differs from the preferred location, update ./docs/spec.md (SDD) before implementation continues.

Record the selection

Only after the joint preflight passes:

• Set the region: azd env set AZURE_LOCATION "<region>" (SDD).
• Declare each deployment in azure.yaml services.ai-project.deployments[] with its deployment name, model name, format (OpenAI, Microsoft, MistralAI, Cohere, DeepSeek, Meta, xAI, …), version, and sku name (usually GlobalStandard) and capacity (thousands of tokens per minute).
• Record the preferred and deployed locations, each deployment and the evidence in .azure/deployment-plan.md.

Workflow

1. Take the task and modality (chat, reasoning, coding, multimodal, embeddings, rerank, image generation, video, audio, document AI) from the spec or brief; ask only when neither states it.
2. Recommend the starting point for the task and resolve its current GA version live.
3. Confirm the preferred region. Start with eastus2 (broadest Azure OpenAI coverage); prefer swedencentral for EU data residency.
4. Confirm capacity. Defaults: 100 (chat / reasoning / coding), 30 (embeddings), 50 (image / audio). When one end-to-end request makes several model calls, size chat capacity to at least four times its estimated tokens, within the remaining quota, so Verify can run E2E cases in parallel. GlobalStandard bills per token, so the headroom costs nothing; record the sizing in the deployment plan.
5. Flag any registration-required or Preview models explicitly before deploying.
6. Run the joint live preflight for the model and every co-located required service; rank only complete placements.
7. Record the selection.
8. Reconcile actual availability. If provisioning selects a different model or location, update ./docs/spec.md and ./.azure/deployment-plan.md immediately.