You train regression models on some rows and score them on rows they never saw, using the fit, predict and score pattern that every aprender estimator shares. You then open model files without running them, reading their tensors, shapes and storage types with apr inspect and apr tensors. You measure what int8 and int4 quantization save in size and cost in answer quality on one question. The module closes by loading a model once with apr serve and talking to it through the OpenAI API shape from curl and from OpenCode.

You turn models into programs. First apr compile builds a downloaded language model into one standalone executable, and include_bytes! compiles a trained model straight into a Rust CLI. You then take a trail-status classifier from raw data to release: look at the data before choosing, compare candidate models on the same features, and save the winner as one small file. You finish by putting that classifier behind an HTTP API that shares one library with the CLI, and by collecting habits that keep inference fast, predictable and easy to call.

You see aprender at work outside the course examples, in an open-source application built on the framework, and look at how its model code, tests and releases fit together. You then find where to go next: the aprender repository, its book, and examples that are tested on every change so they keep working as the framework moves. The module closes the course by gathering what you can now build with Rust and aprender, from a local model answering prompts to a compiled program serving predictions over HTTP.