AI Wisdom

Amazon Web Services associate level

AWS Certified Machine Learning Engineer – Associate

Code
MLA-C02
Length
170 minutes in beta
Pass mark
720 out of 1,000
Price
$75 in beta. MLA-C01 was $150.
Valid for
3 years. Renew by passing the latest version, passing AIP-C01, or through the maintenance option on AWS Skill Builder.
Study time, our estimate
70 to 100 hours

With the ML Specialty retired, this is the AWS exam for ML engineers, and it is mid-change. MLA-C02 adds vector stores, RAG, foundation models, agents and LLM evaluation to classic SageMaker work. The English MLA-C01 ended on September 28, 2026, so English candidates take the C02 beta at $75.

Take it if

You build and operate ML on SageMaker AI and Bedrock.

Skip it if

You only call foundation models. AIP-C01 fits that work better.

Before you start

Check yourself against this list first. If something is missing, close that gap before you pay for an exam.

  • One year with SageMaker AI and BedrockAWS recommends this.
  • One year in a related roleBackend, DevOps, data engineering or data science.
  • Traditional ML and GenAI experienceC02 covers both.

Helpful first: AIF-C01 Next: AIP-C01

All exam details
Format
85 questions in beta. The final exam guide lists 65, with 50 scored.
Length
170 minutes in beta
Pass mark
720 out of 1,000
Price
$75 in beta. MLA-C01 was $150.
Valid for
3 years. Renew by passing the latest version, passing AIP-C01, or through the maintenance option on AWS Skill Builder.
Delivery
Pearson VUE, test center or online
Status
Beta delivery since September 29, 2026. AWS's blog gives general availability as January 14, 2027.
Experience
1 year with SageMaker AI and Bedrock, 1 year in a related role
If you fail
Wait 14 days

What the exam covers

  • Data Preparation for ML and AI28%
  • ML Model and Foundation Model (FM) Development24%
  • Deployment and Orchestration of ML and AI Workflows24%
  • Operating, Monitoring, and Securing ML and AI Solutions24%

What to focus on

  • Added in C02: vector databases, embedding models, chunking, and preparing data for fine-tuning.
  • Added in C02: choosing models in Bedrock, RAG patterns, and NLP metrics such as BLEU, ROUGE and BERTScore.
  • Added in C02: deploying agents, Bedrock Custom Model Import, and reranking.
  • Still core: SageMaker algorithms, feature engineering, bias checks, and hyperparameter tuning.
  • Endpoint choices: real-time, serverless, asynchronous and batch.
  • CI/CD for ML, drift monitoring, cost, and security.
Everything on one mapEvery topic and key idea in this certification.
mindmap
  root((MLA-C02))
    Data preparation
      Vector stores
      RAG prep
      Bias checks
    Model development
      Choosing
      Overfitting
      NLP metrics
    Deployment and orchestration
      Endpoints
      Pipelines
      Agents
    Operate and secure
      Drift
      Cost
      Security

Lesson 1 of 4, 28% of the exam

Data preparation

The largest domain, now with vector and RAG work.

Vector stores
C02 names OpenSearch Service, RDS with pgvector, and Amazon S3.
RAG prep
Chunk documents and extract metadata.
Bias checks
Clarify measures imbalance before training.
Two data paths in C02Classic ML and RAG prepare data differently.
flowchart LR
  A["Raw data"] --> B["Clean and validate"]
  B --> C["Feature engineering"] --> D["Model training"]
  B --> E["Chunk and add metadata"] --> F["Embeddings"] --> G[("Vector store")]

Common trap

Using MLA-C01 notes. C02 added this content.

Try it yourself

Chunk documents into a vector store and run a bias report.

Official docs: SageMaker ClarifyBedrock Knowledge Bases

Lesson 2 of 4, 24% of the exam

Model development

Covers traditional models and foundation models.

Choosing
Weigh custom models, managed services and foundation models.
Overfitting
Regularize or simplify the model.
NLP metrics
C02 names BLEU, ROUGE and BERTScore.
Reading training resultsThe gap between the curves tells you what to do.
flowchart TD
  A{"Training and validation scores"} -->|"Both low"| B["Underfitting: more capacity"]
  A -->|"Train high, validation low"| C["Overfitting: regularize"]
  A -->|"Both high"| D["Ready to test further"]

Common trap

Using classification metrics to judge generated text.

Try it yourself

Tune a model, then evaluate two foundation models.

Official docs: Bedrock evaluations

Lesson 3 of 4, 24% of the exam

Deployment and orchestration

Choosing an endpoint type is a classic question.

Endpoints
Real-time, serverless, asynchronous and batch suit different traffic.
Pipelines
Automate training and deployment with rollback.
Agents
C02 adds agent deployment and state management.
Which SageMaker inference option?Follow the traffic pattern.
flowchart TD
  A{"Do you need an answer now?"} -->|No, all at once| B["Batch transform"]
  A -->|Later, large payload| C["Asynchronous inference"]
  A -->|Yes| D{"Is traffic steady?"}
  D -->|Yes| E["Real-time endpoint"]
  D -->|Sporadic| F["Serverless inference"]

Common trap

Keeping a real-time endpoint on for sporadic traffic.

Try it yourself

Deploy to each endpoint type and note the differences.

Official docs: Deploying models on SageMakerSageMaker Pipelines

Lesson 4 of 4, 24% of the exam

Operate and secure

Models degrade after launch.

Drift
Model Monitor compares live data with a baseline.
Cost
Choose instance families and purchasing options, and watch token use.
Security
Use least privilege, VPC isolation and logging.
Drift detectionMonitoring compares live traffic with a baseline.
sequenceDiagram
  participant E as Endpoint
  participant C as Captured data
  participant M as Model Monitor
  participant A as Alarm
  E->>C: Capture requests
  M->>C: Compare with the baseline
  M->>A: Violation found
  A->>E: Trigger retraining pipeline

Common trap

Scaling up when the problem is drift.

Try it yourself

Set up Model Monitor and a cost alarm.

Official docs: SageMaker Model Monitor

Study plan

  1. Weeks 1 and 2: data

    Prepare a dataset with SageMaker tooling, run bias checks, then chunk documents and load a vector store.

  2. Week 3: models

    Train and tune a model in SageMaker AI. Then evaluate two foundation models in Bedrock on the same task.

  3. Week 4: deploy

    Deploy to each endpoint type and note when each makes sense. Build a pipeline with a manual approval step.

  4. Week 5: operate

    Add Model Monitor, set cost alarms, lock down the endpoint, and deploy one agent.

  5. Exam week

    Read the comparison page in the official guide, then take the mock below.

Exam day and after

  • Unanswered questions count as wrong and there is no penalty for guessing, so answer everything.
  • Scoring is compensatory. You need the overall score, not a pass in each domain.
  • Some questions are unscored and they are not marked, so do not read anything into a strange one.
  • After passing any AWS exam, you get 50% off your next one. A failed attempt means a 14-day wait, and every attempt is paid.
  • Beta results take time. AWS's page says results are typically available within 5 business days of exam completion.
  • AWS's exam table lists the C02 exam code as ME1-C02, so look for that when you book.
  • Japanese, Korean and Simplified Chinese candidates can still take MLA-C01 until C02 reaches general availability.
Test yourself8 practice questions, scored by exam topic.
Start the mock exam

Keep going

Exam facts checked on October 31, 2026 against MLA-C02 official exam guide. Lessons are our own explanations, so use the official docs as the final word, and confirm price and dates before booking.