AI-300 practice test
Microsoft Machine Learning Operations Engineer Associate
100 practice questions, spread across the exam's topics the same way Microsoft weights them, each with a full explanation and a link to the official documentation that proves the answer. Written by a Microsoft Certified Trainer who does this work for a living.
No card needed to start, and none to keep going: buying is not open yet.
What is inside
Every area the exam tests
Our questions are spread the same way Microsoft weights the exam, so your practice matches the real thing.
MLOps infrastructure
15 to 20%Creating and managing workspaces, datastores and compute targets, identity and access for a workspace, data assets, environments and components, sharing assets between workspaces through registries, restricting network access, deploying the platform with Bicep and the CLI, automating provisioning with GitHub Actions, and keeping machine learning projects under source control.
18 questionsModel lifecycle and operations
25 to 30%Experiment tracking with MLflow, automated machine learning, notebooks for exploration, hyperparameter tuning, distributed training, training pipelines and comparing performance across jobs, model registration and versioning, packaging a feature retrieval specification with the model, responsible AI evaluation, archiving, deployment to real-time and batch endpoints, testing and troubleshooting them, progressive rollout and rollback, and detecting drift and triggering retraining.
28 questionsGenAIOps infrastructure
20 to 25%Foundry resources and project environments, managed identities and role-based access, private networking, deployment with Bicep and the CLI, foundation models on serverless endpoints against managed compute, choosing a model for a use case, versioning and production deployment strategies, provisioned throughput for high volume, and holding prompts under version control with variants compared against each other.
24 questionsQuality assurance and observability
10 to 15%Test datasets and data mapping for evaluation, quality metrics including groundedness, relevance, coherence and fluency, risk and safety evaluations for harmful content, automated evaluation workflows with built-in and custom metrics, continuous monitoring, latency and throughput, token consumption and cost, and the logging and tracing that make a production incident diagnosable.
15 questionsOptimising generative systems
10 to 15%Tuning retrieval through similarity thresholds, chunk sizes and retrieval strategy, selecting and fine-tuning embedding models for a domain, combining semantic and keyword retrieval, measuring retrieval quality and testing changes against each other, advanced fine-tuning methods, synthetic data, and taking a fine-tuned model from development into production and watching it there.
15 questionsHow you will use it
Practise the way that suits you
- Learn as you go. Answer, read why straight away, then move on. No clock, no pressure.
- Or sit it timed. A clock, no answers until the end, and two minutes a question, so you learn to pace yourself.
- Score by topic. See which area let you down instead of one number out of a hundred.
- Go back over it. Read every question again afterwards, right and wrong, with the full explanation.
Is it for you
Who takes AI-300
Engineers who put models into production rather than build them
It is about pipelines, registries, endpoints and monitoring rather than about modelling.
Platform teams running an AI capability for several teams
Workspaces, registries, networking and infrastructure as code are examined directly.
Anybody responsible for a generative application after it ships
Two areas are evaluation, observability, cost and the tuning that keeps answers accurate.
Also worth doing
Other exams in the same area
See the questions before you pay
10 AI-300 questions with full explanations, so you can judge the quality yourself.
Start 10 free questions