Prepstellar
Get it on Google Play

AI-103 · Practice set 5 of 13

Choosing LLMs, Small Models, and Multimodal Models: 10 practice questions

10 questions · Untimed · Free

10 free AI-103 practice questions on Choosing LLMs, Small Models, and Multimodal Models, with an explanation for every answer. Untimed. The full mock exam and the timed version are in the app.

Set 5 · Choosing LLMs, Small Models, and Multimodal Models · 10 questions Read the lesson

  1. Question 1 of 10

    Which reasoning effort level does GPT-5 use by default?

    1. AMinimal, the level tuned for throughput
    2. BMedium, the balanced general-purpose level
    3. CLow, the light level for quick triage
    4. DHigh, the level for the hardest problems
    Show the answer

    Reasoning effort trades depth for speed and cost. When you don't set it, GPT-5 uses Medium, the safe middle that suits content drafting, moderate coding, and RAG question answering. You only move away from it on purpose: down for speed, up for hard multistep problems.

    Next → 1 / 10

  2. Question 2 of 10

    A desktop app has downloaded its chat model before a trip. It must then work with no network connection and keep prompts on the user's machine. What fits?

    1. AFoundry Local with a model from its own catalog
    2. BA managed compute deployment in a Foundry hub
    3. CThe model router, to pick the cheapest cloud model
    4. DA serverless deployment of a mini model in Foundry
    Show the answer

    The deciding requirement is where the model runs. Foundry Local processes prompts on the device and works offline after the model files have been downloaded. Every cloud deployment still needs the network and receives the prompt.

    Next → 2 / 10

  3. Question 3 of 10

    Which part of a model card tells you the supported data type of a model?

    1. AThe Benchmarks tab
    2. BThe Deployments tab
    3. CThe Details tab
    4. DLicense tab
    Show the answer

    When a task needs images or other non-text input, check the candidate first. The Details tab carries the description, version information, and supported data type, so it's where you confirm that a multimodal model accepts what your users will send.

    Next → 3 / 10

  4. Question 4 of 10

    A customer service assistant answers live questions from an efficient search index and must feel fast and friendly. Which model suits it?

    1. AGPT-5 at High effort
    2. BGPT-5 at Medium, its default effort
    3. CGPT-4.1, built for low latency
    4. DGPT-5-nano at Minimal with parallel tools
    Show the answer

    Live chat rewards speed over reasoning depth. GPT-4.1 is the non-reasoning, fast response model, built for real-time chat and customer support with low latency. GPT-5 is the reasoning model, and its depth adds latency that a live assistant doesn't need.

    Next → 4 / 10

  5. Question 5 of 10

    An agent built on GPT-5-mini must issue several tool calls in parallel. Which reasoning effort level can't it use?

    1. AHigh
    2. BLow
    3. CMedium
    4. DMinimal
    Show the answer

    Reasoning effort also limits features, not only speed. Parallel tool calls aren't supported at Minimal, so an agent that needs them picks Low, Medium, or High. The rule applies to GPT-5, GPT-5-mini, and GPT-5-nano alike.

    Next → 5 / 10

  6. Keep the ones you got wrong

    In the app, every question you miss comes back exactly when you’re about to forget it.

    Get it on Google Play

  7. Question 6 of 10

    A team wants one server running Foundry Local to answer hundreds of concurrent users. What's the right guidance?

    1. AEnable the optional local web server, which adds continuous batching
    2. BSwitch to the Rust SDK, which handles multi-user GPU sharing in-process
    3. CAdd an Azure subscription so Foundry Local can queue the requests
    4. DUse a dedicated server framework, such as vLLM, for concurrent users
    Show the answer

    Foundry Local is designed for one user on a hardware-constrained device. Server runtimes such as vLLM or Triton Inference Server handle concurrent request queuing, continuous batching, and GPU sharing. For many simultaneous users, pick one of those instead.

    Next → 6 / 10

  8. Question 7 of 10

    Which input and output pairing describes the Claude models in the Foundry catalog?

    1. AText input only, text and image output
    2. BText and image input, image output
    3. CAudio input, text transcript output
    4. DText and image input, text output
    Show the answer

    The Claude family is a multimodal example: it reads images as well as text, with vision and multilingual capabilities, and it answers in text. That's the profile you need when users send photos and expect a written answer.

    Next → 7 / 10

  9. Question 8 of 10

    Analysts want a report that synthesizes about 200K tokens of filings into a multistep argument. They accept a wait of several minutes, but the report must be as accurate as possible. Which setup fits?

    1. AGPT-4.1-nano for high throughput
    2. BGPT-5 at Minimal to cut wait time
    3. CGPT-4.1 in its long-context form
    4. DGPT-5 at High reasoning effort
    Show the answer

    Two constraints decide it: the input fits within the 272K input tokens of GPT-5, and accuracy matters more than speed. A reasoning model at High effort gives the deepest multistep reasoning. The long context of GPT-4.1 is only needed when the input itself is too large.

    Next → 8 / 10

  10. Question 9 of 10

    Field technicians record confidential notes at sites with no connectivity. Their app has cached its model files before travel. Budget rules out per-token charges, and the team has no Azure subscription. Which approach fits?

    1. AA serverless deployment with public network access off
    2. BShip the app with Foundry Local and a curated model
    3. CA GPT-5-nano deployment with Minimal effort for cost
    4. DA managed compute deployment behind a private endpoint
    Show the answer

    No connectivity rules out cloud calls, and the budget rules out per-token cloud pricing. Foundry Local processes notes on the device without an Azure subscription. Its model files must be downloaded before the technicians leave the network.

    Next → 9 / 10

  11. Question 10 of 10

    Users will upload product photos and ask questions about them. Before deploying, the team must prove each shortlisted model accepts images and then compare quality. Which approach fits?

    1. ACheck the License tab for data types, then pick by the Lifecycle filter
    2. BRely on the provider's Responsible AI review, which guarantees fitness
    3. CConfirm image input in each card's Details tab, then use Compare models
    4. DFilter by the Industry facet, then deploy the first text-only model
    Show the answer

    The order matters: first confirm the model can read images, then compare quality. The Details tab lists the supported data type, and Compare models or the leaderboard show benchmark metrics for select models. Skipping the first check risks picking a model that can't read photos.

    Next → 10 / 10

  12. You’ve finished this set

    That’s 10 questions on Choosing LLMs, Small Models, and Multimodal Models. In the app the ones you miss come back exactly when you’re about to forget them.

    Back to the course Next lesson: Foundry Tools or a Model? Get it on Google Play
Get it on Google Play

The whole course, on your phone

Lessons you can read, audio you can listen to on the way to work, and practice that remembers what you got wrong.

Get it on Google Play

“This app is the ‘One Lake’ for preparing the certification exams. Really worth it.”

Abdelhak Y. Google Play review