How Long Does It Take to Prepare for the AWS Machine Learning Engineer Associate (MLA-C02) Exam?

Which exam you sit depends on your test date

The updated exam is in beta, and in English the beta and the outgoing exam do not overlap by a single day. Registration for the beta, exam code ME1-C02, opened on September 1, 2026, and beta delivery starts on September 29, 2026. The last day to take MLA-C01 in English is September 28, 2026. The beta is English only, and the standard MLA-C02 exam is expected in all four exam languages in early 2027.

So the calendar decides: if you are not booked and ready for MLA-C01 by September 28, you are studying for MLA-C02, and everything below is written for that case.

The changeover dates are from the comparison page of the official MLA-C02 exam guide; the beta code and language availability are from the AWS Training and Certification announcement. The outgoing exam's timeline is in the MLA-C01 study-plan post.

Pick your timeline by starting point

These assume 7 to 8 hours a week. The exam guide describes its target candidate as someone with at least one year of experience using Amazon SageMaker AI, Amazon Bedrock, and other AWS services for ML engineering, plus at least one year in a related role (backend developer, DevOps developer, data engineer, or data scientist), with experience in both traditional ML and generative AI. It expects two toolchains, and most adjacent-role candidates arrive with one, which is why MLA-C02 takes longer than MLA-C01.

The target candidate description is from the official MLA-C02 exam guide. It describes who the exam is written for, not an entry requirement. The timelines are our own estimates.

Four domains, almost even, and every one now carries AI

MLA-C02 keeps four content domains and spreads its 50 scored questions across them at 28%, 24%, 24%, and 24%. The largest domain is four points ahead of the other three, about two scored questions.

Domain titles and weights from the official MLA-C02 exam guide, applied to 50 scored questions and rounded.

Two things changed from MLA-C01, which ran 28 / 26 / 22 / 24: model development gave two points to deployment and orchestration, and AWS added generative AI, agentic AI, and foundation model workloads without adding a domain. The guide's comparison page maps all twelve MLA-C01 task statements onto the twelve MLA-C02 tasks one to one, so the new material is new skills inside tasks you already knew. That is the planning headline: there is no separate GenAI week to bolt onto an MLA-C01 plan, because the additions sit in every domain you were already studying.

What the update actually adds

MLA-C01 validated implementing, deploying, and maintaining ML solutions. MLA-C02 validates building, operationalizing, deploying, and maintaining AI and ML solutions, for traditional models and foundation models alike. The comparison page lists the added skills task by task; these are the ones that reshape a plan:

• Domain 1: vector databases (OpenSearch Service, RDS with pgvector), embedding models, chunking and metadata extraction for RAG, masking and anonymizing data, and preparing data for FM fine-tuning, continued pre-training, and distillation.
• Domain 2: selecting FMs from Amazon Bedrock, fine-tuning strategies, RAG architecture patterns, prompt engineering versus fine-tuning, and evaluating generative output with NLP metrics (BLEU, ROUGE, BERTScore), human-in-the-loop review, and LLM-as-a-judge.
• Domain 3: FM deployment options, Bedrock Custom Model Import, deploying and configuring agents, Bedrock knowledge bases, agent state management, GPU scaling, Prompt Management, and pipelines for agent versions and knowledge base refreshes.
• Domain 4: CloudWatch generative AI observability and Bedrock evaluations, agent coordination and tool failures, FM inference and token costs, Bedrock API keys versus IAM credentials, and Bedrock Guardrails.

A few MLA-C01 items were removed: loading training data from EFS and FSx, model size reduction, edge optimization with SageMaker Neo, bring your own container, and two infrastructure monitoring statements. The format keeps its shape: 65 questions, 50 scored and 15 unscored, a passing score of 720, and compensatory scoring, one overall total with no per-domain minimum.

Additions and deletions from the Comparison of MLA-C01 and MLA-C02 page of the official exam guide. Format and scoring from the guide itself.

A concrete 11-week plan (the typical case)

Already MLA-C01 ready? Close the gap in 3 to 5 weeks

If the cutoff caught you, the weights say where the extra weeks go. Deployment and orchestration gained the two points model development lost, and the new material (RAG deployment, agent orchestration, Bedrock at scale) lands hardest there. Start with it, not with prompt theory, then add the Week 10 build and a full drill if you have the time.

• Week 1: Domain 3. FM deployment options, Bedrock knowledge bases, agents, and the agent and knowledge base pipelines, on top of the endpoint and CI/CD material you already know.
• Week 2: Domain 2. Selecting FMs from Bedrock, fine-tuning strategies, RAG patterns, and the generative evaluation toolkit: NLP metrics, human review, LLM-as-a-judge.
• Week 3: Domains 4 and 1. Generative AI observability, agent and token costs, Guardrails and credential types, then vector databases, embeddings, and chunking for RAG. Drop SageMaker Neo, bring your own container, and model size reduction, which MLA-C02 removed.

The trap is that the additions are spread across all four domains rather than parked in one. Skip them and you lose two or three questions in each domain, which looks like noise on a practice exam until it adds up to the gap between 700 and 720.

What makes it take longer

• Two toolchains, not one. The target candidate has a year with SageMaker AI and Bedrock. Most candidates have depth in one and vocabulary in the other.
• It is a beta with a new guide. The exam guide is the only settled scope document, and material written for MLA-C01 does not cover the additions.
• There is still no domain to skim. At 28 / 24 / 24 / 24 the smallest domain is worth about 12 scored questions, and compensatory scoring will not rescue a domain you never studied.

Coming from a Databricks background instead? The Databricks Generative AI Engineer Associate timeline is the closest reference point for how the generative half of this exam will feel.

The cheapest hours in the plan

Reading the task statements, the comparison page, and the in-scope services list in the MLA-C02 exam guide takes an evening and is free. Questions are written from those statements, the comparison page lists every added and removed skill, and the in-scope list is where you learn that Amazon Bedrock AgentCore is on it. For a beta with no settled third-party material, it is the blind-spot audit that matters most.

FAQ