Raschka released an 11-hour video course on building reasoning models in PyTorch

⚡ The gist in 5 seconds

• Sebastian Raschka has published 8 videos and over 11 hours of code on building a reasoning model from scratch in Python and PyTorch.
• The starting point is a small pretrained Qwen3; early examples run on CPU, while training requires a GPU.
• Basic Python and PyTorch are needed; the previous LLM course is recommended but not a mandatory prerequisite.

🔍 What happened

On October 24, Sebastian Raschka, PhD, published a full video course in Ahead of AI titled "Building Reasoning Models from the Ground Up": eight videos, 11+ hours of coding, covering topics like model evaluation, inference-time scaling, reinforcement learning, and distillation. The material pairs with his book Build A Reasoning Model (From Scratch) and continues last year's course Coding LLMs from the Ground Up. On Hacker News, the post garnered 598 points and 85 comments.

The mechanics are simple and all the more pleasant for it: you don't start with an empty file, but take a ready-made small Qwen3 and fine-tune it by hand. First — text generation and evaluation metrics, then inference-time scaling, followed by RL with verifiable rewards and distillation. As for hardware: the early chapters run fine on a CPU, while training experiments require a GPU. In the text, Raschka also provides a conceptual framework: a reasoning model is a regular LLM plus post-training via RL with verifiable rewards, optionally with inference-time scaling. Then comes an engine analogy: the model itself is the motor, while the agent harness adds context management and access to tools. The full announcement and links to the videos are in the Ahead of AI article.

🧠 Our take

The course arrives at a time when explaining reasoning models has become harder than using them. Almost any current LLM is either a reasoning model itself or has a thinking mode — as the author himself notes. We believe the deficit has shifted: previously, there was a lack of access to strong models; now, there is a lack of people who understand what the "reasoning effort" knob actually turns and why the token bill depends on it.

The practical value here lies in the scale. 11 hours of code on a small Qwen3 is not about how to beat the frontier, but about how the RL loop with verifiable rewards works: where the reward is computed, what sampling looks like, and why distillation is sometimes cheaper than direct training. These things are usually explained on slides, not in an editor. From experience, it is precisely building the pipeline by hand that best dispels the illusion that "the model just thinks longer."

Who wins: students, engineers without an ML background, and small teams that need to explain to a client what GPU hours are paying for. Who loses: those selling the "magic" around reasoning. Our assessment is that the course hits infoproducts that promise the same thing for money, without the ability to run the code. There are no losers in terms of technology here: the clearer the mechanics, the less noise in the market.

⚠️ What we don't know

The announcement contains neither the parameter count of the Qwen3 used, nor VRAM requirements, nor how long a single training run takes. It doesn't say which datasets the RL experiments are built on or whether the results converge on a weak GPU. Separately: the date at the top of the article is October 24, 2026, which is odd, and we won't claim whether it's a typo or not. The links to the videos themselves are cut off in the provided fragment — open the original source and check the playlist contents yourself.

Prediction: within a quarter, forks of the repository will appear around the course, along with at least one Russian translation; some forks will use other base models instead of Qwen3.

Source: Ahead of AI (Sebastian Raschka)