October 27, 2026 — Noon

Eleven Hours of Building Reasoning Models From Scratch

Sebastian Raschka just released a full video course on building reasoning models from the ground up in Python and PyTorch. Eight videos, over 11 hours of content, starting from a pre-trained Qwen3 model and implementing evaluation, inference-time scaling, reinforcement learning with verifiable rewards, and distillation from scratch.

This is essentially the sequel to his "Build a Large Language Model (From Scratch)" course from last year. The new course stands on its own, though comfort with Python and PyTorch is recommended.

What reasoning models actually are

Raschka frames it cleanly: a reasoning model is a conventional LLM upgraded with additional post-training via reinforcement learning with verifiable rewards, plus optional inference-time scaling techniques. The goal is making the model better at harder problems, classically math and code, though the capabilities generalize.

The reasoning model then serves as the engine inside an agent harness. The harness handles context management and tool access. The model handles the thinking. Most current-generation LLMs are now reasoning models with variable reasoning effort settings.

"A reasoning model is essentially an upgrade from a conventional LLM, adding more post-training via reinforcement learning with verifiable rewards (and optionally inference scaling techniques) to make it more capable of solving more challenging problems."

The curriculum

The course covers four major areas:

• Evaluation — measuring what a model can actually do, not what it looks like it can do
• Inference-time scaling — spending more compute at inference to get better answers, the technique that made models like GLM-5.2 and o1 dramatic step changes
• Reinforcement learning with verifiable rewards — training on problems where answers can be checked automatically (math, code), which is why reasoning models emerged from those domains first
• Distillation — compressing a larger model's reasoning ability into a smaller, faster one

The early text-generation and evaluation examples can run on a CPU. The later training experiments need a GPU.

Why this matters

Reasoning models are now the default. Almost every major LLM release in 2026 includes a reasoning mode or is a reasoning model outright. But the mechanics of how they work, the RL training loop, the inference-time compute tradeoff, the distillation pipeline, these are often treated as black boxes even by people building on top of them.

Raschka's approach is the opposite: implement it yourself, understand every component, then use the production tools with actual knowledge of what they do. His "Build an LLM From Scratch" book was one of the clearest resources for understanding transformer internals. This reasoning course applies the same philosophy to the post-training side.

Companion book

The videos accompany his book "Build a Reasoning Model (From Scratch)." The recommendation is to read a chapter, work through the code alongside the corresponding video, then try the exercises. The book provides the depth, the videos provide the walkthrough.

The entire course is free. The book is the paid companion if you want exercises and deeper written explanations.

If you have been treating reasoning models as magic boxes that sometimes think for too long and sometimes not enough, this course is the antidote. 11 hours is a serious commitment, but it is shorter than the time most people spend guessing at what inference-time scaling actually does.

Sources:

[1] Sebastian Raschka, "Building Reasoning Models from the Ground Up: A Full Video Course," Ahead of AI, October 24, 2026. https://magazine.sebastianraschka.com/p/building-reasoning-models