LLM/ML Courses Roundup: 14 Picks, 3 Tracks (2026)

If you want to actually understand large language models rather than just call an API, this LLM/ML courses roundup is where to start. Nineteen courses came across the desk, fourteen made the cut, and they fall into three honest tracks: foundations, retrieval, and agent engineering. Every pick links to a free, public course from a primary source, with a one-line take on what you get and when to use it. Nothing here is sponsored, and every course can be started without paying; where certificates exist they cost extra, the lessons do not.

The library is organized by what you are trying to build, not by vendor. Start at the bottom of the stack (how a transformer actually works) and climb up (retrieval, then autonomous agents). Three tracks, one path.

How the fourteen were screened matters as much as the count. Every pick had to be free to start and taught by people close to the software in the exercises. The agent memory course comes from the founders of Letta, the team behind MemGPT. Both crewAI courses are taught by the founder of crewAI. The advanced RAG course is taught by Jerry Liu and Anupam Datta together with LlamaIndex and TruEra, the tools its labs run on. When the instructor ships the tool you practice with, the exercises stay honest, because they come from real design decisions rather than vendor demos.

Foundations: build from scratch first

Two Karpathy courses anchor the foundations track. Neural Networks: Zero to Hero is eight lectures that start at backprop internals and end with you training a GPT by hand: micrograd first, then the makemore series through MLPs, batch normalization, and WaveNet, then the GPT build, with a tokenizer lecture to close. Language Modeling from Scratch (the nanoGPT lecture) is the capstone. The point: you cannot debug what you cannot see, and these courses make the internals visible. Basic Python is the only real prerequisite. See the full foundations roundup for the two entries plus our inference foundations cross-link.

RAG and vector databases: ground the model in your data

Retrieval-augmented generation is where most production LLM work actually lives. This track covers the full stack: vector storage with Pinecone (Tim Tully builds six applications on it, from semantic search to anomaly detection), embeddings-to-app pipelines with Sebastian Witalec, advanced retrieval and its metrics from Jerry Liu and Anupam Datta, and the accuracy course where Sharon Zhou and Amit Sangani turn a working demo into a dependable one. The fifth pick is different in kind: a 26-hour intermediate build of the whole pipeline with Zain Hasan, graded assignments included. Five courses, each covering a distinct layer. See the RAG and vector databases roundup.

Agent engineering: memory, tools, and multi-agent systems

Agents are the 2026 frontier and the most overhyped layer. This track cuts the noise: agent memory from the Letta founders, browser agents (Div Garg and Naman Garg) and computer use (Colt Steele, with Anthropic), agent evaluation from John Gilhuly and Aman Khan, the two crewAI orchestration courses taught by the framework’s own founder, and LLMOps (the production layer most tutorials skip). Six courses plus an LLMOps section. See the agent engineering roundup.

Already covered: cross-link only

Five courses from the original list already have dedicated tutorial posts in our AI Agents series: HuggingFace Agents, MCP with Anthropic, Agent Design Patterns, Berkeley Agent MOOC, and Berkeley Advanced Agents. Read those for depth; this hub points you to them rather than repeating them.

New here and unsure where to start? Do Karpathy first (foundations), then one RAG course (Building and Evaluating RAG apps is the fastest to apply), then one agent course (Agent Memory is the clearest). Three courses, roughly forty hours, and you will understand the full stack well enough to debug production issues.

Already shipping? Work backwards from the pain. LLMOps if deploys hurt, agent evaluation if you cannot tell whether a change helped, the accuracy course if the demo works until a real user touches it. The roundups carry the per-course detail: level, duration, and the specific angle that separates each pick from its neighbors.