CertPrep Coach

AI models and configuration

How generative models produce text, how to choose a model, and what deployment types and generation settings control.

How a generative model produces a response A prompt is split into tokens, and the model predicts the response one token at a time. Settings such as temperature and maximum tokens shape the output. The deployment type decides where requests are processed and how you pay. shape output where it runs Prompt system message + user message Tokens text split into small pieces Model predicts the next token, then repeats Response text, code or JSON Generation settings temperature, top_p, max tokens, stop sequences Deployment type Global, Data Zone, Standard; pay per token, provisioned or batch
A prompt is split into tokens, and the model predicts the response one token at a time. Settings such as temperature and maximum tokens shape the output. The deployment type decides where requests are processed and how you pay.

What Microsoft lists for this area

  • Describe how generative AI models work
  • Identify an appropriate AI model, based on capabilities
  • Identify appropriate model deployment options and configuration parameters

What this area tests

Three things: a basic understanding of how generative AI models work, choosing a suitable model for a task, and knowing which deployment options and configuration parameters to use.

How a generative model works

A generative language model reads your prompt as tokens, which are small pieces of text such as words or parts of words. It then predicts the most likely next token, adds it to the response, and repeats until the response is complete. That’s why responses stream in piece by piece, and why limits are measured in tokens rather than words.

A prompt usually has two parts. The system message sets the model’s role, rules and tone. The user message is the actual request.

Choosing a model

The Foundry model catalog offers thousands of models from Microsoft, OpenAI, Meta and others. Choose by capability first:

  • Chat or reasoning models for answering questions, writing and summarizing.
  • Multimodal models when the input includes images or audio.
  • Image generation models to create images.
  • Embedding models to turn text into vectors for search and retrieval.

Then weigh cost, speed and quality. A smaller model is often fast and cheap enough for simple tasks.

Generation settings

  • Temperature: higher values make output more random and creative; lower values make it more focused and consistent.
  • Top_p: an alternative way to control randomness. Change temperature or top_p, not both.
  • Max tokens: caps the length of the response.
  • Stop sequences: text that tells the model where to stop.

If answers to the same prompt vary too much, lower the temperature. If they’re cut off, raise max tokens.

Deployment types

You deploy a model before you can call it. The deployment type decides where requests are processed and how you pay:

  • Global Standard: routes traffic to available datacenters worldwide, billed per token.
  • Data Zone Standard: keeps processing within a Microsoft-defined data zone (US, EU or APAC), billed per token.
  • Standard: keeps processing within an Azure geography, billed per token.
  • Provisioned (global, data zone or regional): you reserve throughput in advance for predictable, high-volume workloads.
  • Batch: processes large jobs asynchronously at a discount, within about 24 hours.

Exam tip: if a scenario mentions data residency, look at Data Zone or Standard. If it mentions steady, high-volume traffic with predictable cost, look at Provisioned.

Practice AI models and configuration

20 practice questions with explanations for this area.

Practice this area Or read 12 sample questions with answers

Sources

Updated October 25, 2026. Microsoft changes products often; check the official study guide for the latest outline.