Plan and manage an Azure AI solution 1.2 Set up AI solutions in Foundry
A team wants to deploy a Qwen model from the Hugging Face collection in the Foundry model…
Study guide skill: Choose appropriate deployment options
Question B6-006
A team wants to deploy a Qwen model from the Hugging Face collection in the Foundry model catalog. It should run on dedicated GPU capacity, and the team doesn't want to manage virtual machines, clusters, or serving runtimes. Which deployment option should they use?
Select one answer.
Answer: B
Managed compute (preview) hosts open-source and custom-weight models, including models from the Hugging Face collection such as Qwen, on dedicated GPU capacity, with no virtual machines, clusters, or serving runtimes to own. Instant access (preview) isn't a deployment option; it calls supported models by name without creating a deployment. Developer and Batch are deployment types of the serverless API option, which is for Foundry Models.
Deep dive
How it works
Foundry has two deployment options:
- Serverless API: for Foundry Models, including Foundry Models sold by Azure and select models from partners and community. Within it you pick a deployment type (Global Standard, Data Zone Standard, Standard, provisioned, batch, developer and more).
- Managed compute (preview): for open-source and custom-weight models. You can deploy models from the Hugging Face collection this way, and the page lists Qwen models among the examples.
Managed compute is built for exactly this team's constraint. You reach the deployment through the same Foundry project endpoint as other deployment types, with no virtual machines, clusters or serving runtimes to own. It's a PaaS offering, not IaaS. Foundry chooses GPUs per instance from model size, architecture, context length and whether you optimize for latency or throughput, so you don't pick VM SKUs. Billing is hourly per accelerator SKU, and quota is granted per accelerator SKU per region, separate from Azure VM quota.
The other options, and when you'd use them
- Serverless API with the Developer type: Developer is one of the serverless deployment types, and serverless is for Foundry Models. Hugging Face collection models go through managed compute.
- Serverless API with the Batch type: Same issue. Batch is another serverless deployment type, not a way to run a Hugging Face model on dedicated GPUs.
- Instant access (preview): It isn't a deployment option at all. It lets you call supported models by name without creating a serverless API or managed compute deployment.
Exam tip
Hugging Face / open-weight model on dedicated GPUs without owning infrastructure = managed compute (preview). Developer, Batch, Global Standard and so on are serverless deployment types.
References
Last reviewed: