Plan and manage an Azure AI solution 1.2 Set up AI solutions in Foundry
A team has fine-tuned five candidate models and wants a low-cost way to run evaluations…
Study guide skill: Choose appropriate deployment options
Question B4-006
A team has fine-tuned five candidate models and wants a low-cost way to run evaluations against each one before picking one for production. They don't need an SLA or data residency for this testing. Which deployment type is designed for this?
Select one answer.
Answer: A
The Developer deployment type is designed for fine-tuned model evaluation only: it provides cost-efficient testing of custom models without data residency guarantees or an SLA, and Developer deployments are deleted automatically when their fixed lifetime ends. Provisioned types reserve throughput capacity with hourly billing, which suits steady production traffic rather than short evaluations. Data Zone Standard is a pay-per-token type defined by keeping processing within a data-zone boundary, which this testing doesn't need.
Deep dive
How it works
Foundry deployment types fall into three main categories: standard (pay-per-token), provisioned (reserved capacity) and batch (discounted asynchronous processing). A fourth, the Developer type (SKU name DeveloperTier), exists for one job: fine-tuned model evaluation.
What defines Developer:
- It provides cost-efficient testing of custom models and is billed pay-per-token.
- It doesn't include data residency guarantees or an SLA.
- Developer deployments have a fixed 24-hour lifetime and are deleted automatically when it expires.
The team has five fine-tuned candidates, wants low cost, and explicitly doesn't need an SLA or data residency. That's a word-for-word match for Developer. When a winner is picked, deploy it with a production type.
The other options, and when you'd use them
- Data Zone Standard: A pay-per-token type that routes traffic within a Microsoft-defined data zone (US, EU or APAC). You pick it for data-zone compliance, which this testing doesn't need.
- Regional Provisioned: You specify the throughput you need in PTUs, and the architecture page lists its billing as hourly reserved capacity. That suits steady production traffic, not short evaluation runs.
- Global Provisioned: Reserved processing capacity for predictable throughput with global routing, also billed as hourly reserved capacity.
Exam tip
"Evaluate fine-tuned models cheaply, no SLA needed" = Developer (24-hour lifetime). Provisioned = reserved capacity for production.
References
Last reviewed: