Test your understanding of the Model Catalog, deployment types, and model management with these 15 scenario-based questions.
Your team needs to deploy a Meta Llama 3.1 70B model for a proof-of-concept. You want minimal infrastructure management and pay-per-token pricing. Which deployment type should you choose?
A) Azure OpenAI Standard deployment B) Managed Compute with a dedicated GPU VM C) Serverless API (Models as a Service) D) Azure OpenAI Provisioned (PTU)
A production chatbot is experiencing intermittent HTTP 429 errors during business hours. The deployment is Global Standard with 50K TPM allocated. What should you do first?
A) Switch to a Provisioned (PTU) deployment immediately B) Increase the TPM allocation for the deployment C) Implement retry logic with exponential backoff and increase TPM D) Deploy to a second region and load balance
You are building a customer support classifier that categorizes tickets into 10 categories. The task is straightforward and volume is high (200K requests/day). Which model is the most cost-effective choice?
A) GPT-4o B) GPT-4o-mini C) o3 D) Llama 3.1 405B
Your application uses gpt-4o version 2024-08-06. You see a notification that this version will be retired in 90 days. What is the recommended approach?
A) Wait for auto-upgrade and hope nothing breaks B) Deploy the new version in a staging environment, test with your prompts, then update production C) Switch to a different model family immediately D) Ignore it — Microsoft will keep old versions running indefinitely
A startup wants to use Mistral Large for multilingual customer support. They have no ML infrastructure experience. Which deployment option is best?
A) Managed Compute — they need full control B) Serverless API — zero infrastructure, pay-per-token C) Azure OpenAI Standard — it supports Mistral D) Self-hosted on Azure VMs
Your enterprise processes 500M tokens per month with GPT-4o and needs consistent low latency. Which deployment type is most appropriate?
A) Global Standard (PAYG) B) Standard (PAYG) in a single region C) Provisioned Throughput Units (PTU) D) Serverless API
You need to generate text embeddings for a RAG application. The embedding index has 10 million documents and cost is a concern. Which embedding model should you choose?
A) text-embedding-3-large (3072 dimensions) B) text-embedding-3-small (1536 dimensions) C) GPT-4o with a prompt to generate embeddings D) Cohere Embed v3 (1024 dimensions)
A developer deployed a model with version_upgrade_option set to NoAutoUpgrade. The current version is about to retire. What happens?
A) The deployment automatically upgrades to the latest version B) The deployment continues to work indefinitely with the old version C) The deployment stops working after the retirement date — the developer must manually upgrade D) Azure sends a one-time warning and then auto-upgrades
You have a batch processing job that classifies 1 million documents. The results are not time-sensitive — a 24-hour turnaround is acceptable. How should you optimize costs?
A) Use GPT-4o with Global Standard deployment B) Use GPT-4o-mini with the Batch API for 50% savings C) Deploy Llama 3.1 on Managed Compute D) Use Provisioned (PTU) for guaranteed throughput
Your team wants to compare Phi-4 and GPT-4o-mini for a code generation task. How should you evaluate them before committing to a deployment?
A) Read the benchmark scores on the model cards and choose the one with higher HumanEval B) Deploy both, test with your actual prompts and use cases, compare quality and cost C) Choose whichever is cheaper per token D) Always choose the proprietary model — it is guaranteed to be better
You deployed GPT-4o in East US 2 and are using 80% of your regional quota. A new application team wants to deploy GPT-4o in the same region. What should you do?
A) Tell them to use a different model B) Check available quota, request an increase if needed, and allocate TPM for their deployment C) Move your deployment to another region to free up quota D) Share your deployment with the other team
A content moderation team reports that the Azure OpenAI content filter is blocking legitimate medical content in their healthcare application. What should they do?
A) Remove the content filter entirely B) Create a custom content filter policy with adjusted thresholds for the medical domain C) Switch to an open-source model that has no content filter D) Content filters cannot be modified — redesign the application
You have three GPT-4o deployments: production (200K TPM), staging (50K TPM), and a demo environment (100K TPM) that is rarely used. How can you optimize quota usage?
A) Delete all deployments and recreate them B) Reduce the demo deployment to 10K TPM and reallocate the freed quota to production C) Request more quota from Microsoft D) You cannot change TPM after deployment
An application needs both text chat capabilities and image understanding (analyzing uploaded receipts). Which single model deployment covers both needs?
A) GPT-4o (supports text and image input) B) GPT-4o-mini for text + DALL-E 3 for images C) Whisper for text + GPT-4o for images D) text-embedding-3-large for both text and images
Your company requires that all AI data stays within the European Union. Which deployment configuration meets this requirement?
A) Global Standard — Microsoft handles data residency automatically B) Standard deployment in a European region (e.g., Sweden Central) or Data Zone deployment C) Serverless API with Llama — open models have no data restrictions D) Provisioned (PTU) in any region — PTU guarantees data locality
Serverless API is the right choice for partner models like Meta Llama when you want pay-per-token pricing with no infrastructure to manage. Managed Compute would work but requires managing VMs. Azure OpenAI deployments do not support Llama models.
The immediate fix is implementing proper retry logic (a code-level fix), then increasing TPM allocation to reduce the frequency of 429 errors. Switching to PTU is a bigger decision that should be evaluated based on sustained traffic patterns, not a quick fix.
For high-volume, straightforward classification tasks, GPT-4o-mini provides excellent quality at ~90% lower cost than GPT-4o. The o3 model is overkill for simple classification, and Llama 3.1 405B is expensive to run at scale.
The correct approach is to deploy the new version alongside your existing one, run your test suite with representative prompts, verify output quality, and then swap production. Never rely on auto-upgrade for production workloads.
Serverless API is perfect for teams without ML infrastructure experience. It provides Mistral Large access through a simple API with no VMs to manage. Azure OpenAI Service does not host Mistral models.
At 500M tokens/month with low-latency requirements, PTU provides guaranteed throughput and predictable latency. PAYG deployments at this volume may hit rate limits and offer less latency predictability.
For 10M documents where cost matters, text-embedding-3-small offers a strong balance of quality and cost — ~80% cheaper than text-embedding-3-large. GPT-4o is not an embedding model. Cohere Embed is a valid option but text-embedding-3-small is the most cost-effective Azure-native choice.
With NoAutoUpgrade, Azure will not automatically update the model. After the retirement date, the model version is removed and API calls will fail. The developer must deploy a new version before retirement.
Combining GPT-4o-mini (cheaper model) with the Batch API (50% discount for 24-hour SLA) gives the best cost optimization. The task is not time-sensitive, making batch processing ideal.
Benchmarks are a starting point, but real-world performance depends on your specific use case. Deploy both models, run your actual code generation prompts through them, and compare output quality, latency, and cost.
The standard process is to review current quota utilization, determine if there is room for a new deployment, and request a quota increase from Azure if needed. Quota increases are typically approved within a few business days.
Azure AI Foundry allows custom content filter policies where you can adjust severity thresholds per category. For healthcare, you can raise the threshold for medical content while keeping other safety filters active. Filters cannot be fully removed.
TPM allocations can be adjusted at any time for PAYG deployments. Reducing the rarely-used demo environment frees quota for production use. This is a free, immediate change.
GPT-4o is multimodal — it accepts both text and image inputs in a single API call. You can send receipt images along with text instructions. DALL-E generates images but does not understand them.
Standard deployments are regional — data stays in the selected region. Data Zone deployments ensure data stays within a geographic zone (e.g., EU). Global Standard routes data globally, which may not meet EU data residency requirements.