Sebastian Raschka has done this before. His Build a Large Language Model (From Scratch) book became the default recommendation for engineers who wanted to understand transformers by implementing them, not by reading about attention diagrams. This sequel applies the same method — ground-up PyTorch implementation — to the post-training stack that turns a base LLM into a reasoning model: reinforcement learning with verifiable rewards, inference-time scaling, self-consistency, distillation. The course is eight videos totaling 11 hours, built around a pre-trained Qwen3 0.6B model. Starting from a working checkpoint rather than random weights is the right call — it lets Raschka skip tokenizer training and pretraining logistics and focus entirely on the reasoning stack. You load the model, implement KV caching and torch.compile for efficient generation, build a math-answer verifier against MATH-500, then progressively add temperature scaling, multinomial sampling, top-p filtering, self-consistency via majority voting, log-probability scoring, and self-refinement loops. The pedagogical structure is disciplined. Each video maps to a book chapter, and the recommended workflow — read the chapter, code along with the video, do the exercises — reflects Raschka's conviction that implementation is comprehension. This is not a MOOC with quizzes. It is a workshop where you build the thing and measure whether it works. The MATH-500 verifier doubles as both an evaluation tool and the reward signal for later RLVR training, which is a clean design choice that shows how the pieces connect. The scope is deliberately narrow. This is not a survey of reasoning approaches or an introduction to agents. Raschka explicitly positions the reasoning model as the "engine" inside an agent harness and then stays focused on the engine. The agent harness — tool use, context management, orchestration — is mentioned but deferred to a separate Substack article. If you want to understand why your reasoning model produces better chain-of-thought than your base model, this course shows you the mechanism. If you want to build an agent, you need this plus other materials. Hardware requirements are modest for the early videos (CPU-viable) but ramp up for training. A GPU is described as practical rather than mandatory, and the repo includes environment recommendations. The choice of Qwen3 0.6B keeps compute accessible — you are not fine-tuning a 70B model — while still demonstrating real capability differences on MATH-500. The companion book is available from Manning and Amazon in color print. Raschka also offers a 90-minute conceptual overview on LinkedIn Learning for those who want the ideas without the implementation. The full course is free on YouTube. This is a generous release model: the deep material is open, the book is priced, and the LinkedIn course serves as an on-ramp. What you will not find here is novel research or frontier-scale results. This is pedagogy, not a paper. The value proposition is that after 11 hours you will have implemented the core reasoning stack yourself, understand what RLVR actually does to a model's behavior, and be able to read the next DeepSeek or OpenAI technical report with real comprehension of the training pipeline described. For the target audience — engineers who learn by building — that is a substantial offer.
