Sebastian Raschka teaches how to code reasoning models from scratch

Summary

Sebastian Raschka announced a PyTorch course consisting of 8 videos and 11 hours, accompanying his book with Manning. The course walks through how to build reasoning capabilities on top of the pre-trained Qwen3 0.6B model step by step: evaluation, inference-time scaling, reinforcement learning with verifiable rewards, and distillation.

Why it matters

Since nearly all frontier LLMs now ship with a 'reasoning mode,' understanding post-training mechanisms (RLVR, test-time compute, distillation) is critical for practitioners looking to go beyond API calls. Raschka’s from-scratch, code-focused pedagogy has a proven track record of making LLM internals accessible. Pairing this approach with a companion book lowers the barrier to entry for engineers who find paper explanations unhelpful.

Highlights

• The course uses a small Qwen3 0.6B base model to shift focus from pre-training scale to the reasoning post-training pipeline.
• Raschka positions reasoning models as the 'engine' within an agent harness, clarifying a distinction that many practitioners still confuse.
• The series is explicitly standalone: prior completion of the 'Build an LLM from Scratch' course is not required, though it is recommended.

Critical take

The post reads more like a product launch than a technical contribution. Beyond a single paragraph using the engine and harness analogy, it offers no new insights, benchmarks, or code snippets; it merely lists the contents of a paid book and video set. For readers already following Raschka’s free 'Ahead of AI' articles, the added value of this announcement is limited. Additionally, the 'from scratch' label is somewhat generous given that the base model is downloaded as pre-trained weights.