Video course teaches reasoning-model internals with PyTorch implementations

Sebastian Raschka released an eight-part video course totaling over 11 hours of content focused on building reasoning models from scratch using Python and PyTorch. The series covers evaluation, inference-time scaling, reinforcement learning, and distillation, and is designed to complement Raschka’s prior work on large language models.

The course begins with a pre-trained Qwen3 0.6B model and walks through implementing text generation with KV caching, evaluation and reinforcement-learning verifiers, inference-time scaling, and distillation. Early lessons can be run on CPU, while later experiments are more practical with a GPU.

Raschka positions reasoning models as upgrades over conventional LLMs, achieved via post-training such as reinforcement learning with verifiable rewards and optionally inference-time scaling. He notes that most current-generation LLMs are reasoning models with variable reasoning-effort settings and that reasoning models often serve as the engine inside agent harnesses for complex tasks.

The videos are accompanied by a companion book, Build a Reasoning Model (From Scratch), available in print and digital editions. Raschka recommends pairing video lessons with the book’s chapters and exercises for a deeper learning experience. He also highlights a shorter, conceptual LinkedIn Learning course on reasoning models as optional pre-work.