Sebastian Raschka's from-scratch video course on building reasoning models

New video lectures let you build and deeply understand LLM reasoning models yourself, not just use them.

Sebastian Raschka has released eight new coding videos, totaling 11 hours, that guide viewers through building reasoning models from scratch using PyTorch. The series, presented as a sequel to his "Coding LLMs from the Ground Up" course, starts with a pre-trained LLM and focuses on implementing methods for evaluation, inference scaling, reinforcement learning, and distillation. Key topics include setting up the coding environment, loading and generating text from a base Qwen3 model, building a verifier for evaluation and reinforcement learning, and exploring inference-time scaling techniques like sampling, self-consistency, and self-refinement. The course material is designed to be standalone but is best experienced alongside Raschka's accompanying book, "Build a Reasoning Model (From Scratch)." Prerequisites include comfort with Python and PyTorch, with a GPU recommended for later training experiments. The videos aim to provide a deep understanding of how reasoning models function as the engine within modern AI agents.

This release offers a practical, hands-on approach to understanding the core mechanics of reasoning models, positioning them as crucial upgrades to conventional LLMs. By implementing techniques like reinforcement learning with verifiable rewards and inference scaling from the ground up, Raschka demystifies how these models achieve enhanced problem-solving capabilities, particularly in complex domains like math and code. The series bridges the gap between theoretical understanding and practical application, enabling individuals to grasp the 'engine' that powers advanced AI agents.

The availability of this extensive course, along with its companion book, signals a growing trend towards accessible, in-depth education in advanced AI development. It empowers developers and researchers to move beyond black-box solutions and gain granular control over model behavior. This focus on building from scratch is essential for anyone looking to innovate or fine-tune AI systems for specific, challenging tasks, as it provides the foundational knowledge needed to adapt and improve upon existing architectures.