TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Sebastian Raschka has published eight from-scratch coding videos, totaling 11 hours, to accompany his book Build a Reasoning Model (From Scratch). The material uses a pre-trained Qwen3 model to teach evaluation, inference-time scaling, reinforcement learning and distillation in Python and PyTorch.
Computer scientist and author Sebastian Raschka has published eight coding videos totaling 11 hours that walk through building reasoning-model methods from scratch in Python and PyTorch. The lectures accompany his book, Build a Reasoning Model (From Scratch), and use a small pre-trained Qwen3 model to demonstrate evaluation, inference-time scaling, reinforcement learning and distillation.
Raschka describes the videos as a practical, implementation-focused course rather than a requirement to train a language model from the beginning. The material starts with a pre-trained model, then builds and tests methods intended to evaluate and improve its performance. In his Ahead of AI article, he says this approach lets learners focus on the reasoning-related techniques themselves.
The course is designed to stand alone, though Raschka recommends familiarity with Python and PyTorch. His earlier Build a Large Language Model (From Scratch) material may help, but is not a prerequisite. According to the article, early text-generation and evaluation examples can be run on a CPU; a GPU makes later experiments, particularly training, more practical. Raschka says the first video covers setup, with further hardware and environment guidance in the accompanying repository.
The eight lectures accompany the book, which Raschka says is available through Manning and Amazon. He recommends pairing the reading with the videos and exercises. He also points readers to a separate, roughly 90-minute conceptual course on LinkedIn Learning for a shorter introduction to inference-time scaling, reinforcement learning and distillation. That course is presented as an optional overview, not a substitute for the coding series.
From Model Concepts to Working Code
The course offers learners a way to examine the techniques behind reasoning models through code they can follow and run, rather than only through high-level descriptions. That matters as reasoning capabilities and adjustable reasoning modes have become common features of current-generation language models, according to Raschka. Understanding how evaluation, additional inference computation and post-training fit together can help developers and students interpret what those systems do.
Raschka also draws a boundary between a reasoning model and an AI agent. In his explanation, the model is the reasoning engine; an agent harness supplies surrounding software, context management and tool access to handle larger tasks. The distinction helps readers understand why improving a model and building a capable agent are related but separate engineering problems. The videos focus on the model component, not on constructing a complete agent system.
The practical value will depend on a learner’s background, available hardware and how closely the examples match their goals. The announcement describes educational material; it does not provide independent evaluations of the course or evidence that the techniques will produce a particular performance gain on other models.
Products Worth Considering
Bose QuietComfort True Wireless Earbuds with Active Noise Cancellation, Black
Anker SOLIX C200 DC Portable Power Station, 192Wh, 200W
DJI Osmo Mobile 8 Gimbal with Mic Mini Transmitter, Infinity Black
As an affiliate, we earn on qualifying purchases.
How the Course Fits Raschka’s Work
Raschka says he has written about reasoning models on Ahead of AI over the past two years, including a recent article on controlling reasoning effort in contemporary language models. He presents the new video series as a sequel to his earlier Coding LLMs from the Ground Up course, while stressing that learners can start with the reasoning material without completing that earlier course.
The accompanying book is titled Build a Reasoning Model (From Scratch). The videos translate its subject matter into lectures with coding demonstrations. Raschka recommends reading a chapter, working through its code alongside the corresponding video, and then attempting the exercises. He also links to his article on the components of a coding agent as related reading about how models are used within broader systems.
In Raschka’s framing, a reasoning model is a conventional language model further developed through post-training, often using reinforcement learning with verifiable rewards, and sometimes inference-time scaling. He says these models are classically developed for mathematics and code, while their capabilities may extend to other problem areas. These are explanatory claims in his article, not results independently established by the course announcement.
“I wrote my recent Build Reasoning Model (From Scratch), which implements these from the ground up in PyTorch.”
— Sebastian Raschka, in Ahead of AI
Access, Hardware and Course Scope
The supplied article does not specify a publication date for the video series, whether access is free or paid, or whether the videos carry a certificate. It also does not provide a detailed breakdown of the runtime for each lecture. Those points should not be inferred from the announcement.
The source describes the repository as containing environment and hardware recommendations, but the material supplied here does not list specific GPU models, memory requirements or expected training times. Raschka says a CPU is sufficient for early examples and a GPU makes later training experiments more practical; the exact resources required will depend on the exercises and setup.
It is also unclear how closely the implementation transfers to larger or differently designed models. The videos use a small pre-trained Qwen3 model, and the announcement does not report comparative benchmarks, independent reviews or measured performance improvements. The course is presented as a learning resource, not as proof that its methods will improve every model or application.
Learners Can Follow the Eight Lectures
Readers can work through the eight published lectures alongside the book and its exercises. Raschka says the opening video introduces the relationship between conventional language models, reasoning models and agents, then covers the coding environment and setup. The series proceeds into the technical work, including evaluation, inference-time scaling, reinforcement learning and distillation.
Those interested in a shorter introduction can use Raschka’s separate LinkedIn Learning course, described in his article as about 90 minutes. Raschka also says he plans to return to Ahead of AI with more research roundups the following month after a short break. The article does not announce a schedule for additional course lectures or updates.
Key Questions
What did Sebastian Raschka release?
He published eight coding videos totaling 11 hours, covering how to implement reasoning-model methods from scratch in Python and PyTorch.
Which topics do the videos cover?
The series covers evaluation and improvement of a pre-trained model, including inference-time scaling, reinforcement learning and distillation. It uses a small pre-trained Qwen3 model for the demonstrations.
Do learners need to take Raschka’s earlier LLM course first?
No. Raschka says the reasoning-model material is standalone. He recommends Python and PyTorch experience, while describing his earlier from-scratch LLM course as helpful but not required.
What computer is needed to follow the course?
Raschka says early text-generation and evaluation examples can run on a CPU. A GPU is more practical for later experiments, especially training. The supplied announcement does not give specific hardware specifications.
How are reasoning models different from AI agents?
In Raschka’s explanation, the reasoning model is the system’s engine. An agent harness adds software for context management and tool access. The video series focuses on developing the model, rather than building a complete agent harness.
Source: magazine.sebastianraschka.com
