Build Your Own Reasoning Model: A Full Video Course

Overview of Reasoning Model Development

Sebastian Raschka has released a comprehensive video course titled Build Your Own Reasoning Model, providing an 11-hour, eight-part series on the mechanics of modern reasoning-capable LLMs. The course focuses on the transition from conventional LLMs to reasoning models through techniques such as reinforcement learning with verifiable rewards (RLVR), inference-time scaling, and distillation. The curriculum is designed to be accessible to those with a foundational understanding of Python and PyTorch, serving as a sequel to his previous work on coding LLMs from the ground up.

Core Concepts and Methodology

The course emphasizes a "from-scratch" implementation approach, utilizing a pre-trained Qwen3 0.6B model as a base. By starting with a pre-trained model, the curriculum focuses on the post-training and inference-time techniques that define modern reasoning capabilities:

  • Inference Scaling: Implementation of temperature scaling, multinomial sampling, and self-consistency to improve output quality through additional computation.
  • Reinforcement Learning with Verifiable Rewards (RLVR): A deep dive into training loops where models are rewarded for correct mathematical or logical outputs, moving beyond standard supervised fine-tuning.
  • Distillation: Techniques for transferring reasoning capabilities from larger teacher models to smaller, more efficient student models.
  • Evaluation: Building custom verifiers and using benchmarks like MATH-500 to measure model performance and training stability.

The Role of Reasoning Models in AI Agents

Raschka frames the reasoning model as the "engine" within a broader agentic architecture. While a conventional LLM provides basic text generation, a reasoning model adds post-training depth to solve complex, multi-step problems. An "agent harness" then wraps this engine to manage context, tool access, and task orchestration. This distinction clarifies the relationship between model intelligence and the software frameworks that enable autonomous agent behavior.

Educational Philosophy and Community Reception

The course has been met with significant interest from the developer community, with many users highlighting Raschka's pedagogical approach to complex AI topics. Proponents of the "from-scratch" methodology argue that building these systems at a lower level of abstraction is essential for understanding the limitations and behaviors of the models used in production environments.

"Building something at a level further down the stack than you will ever go forces you to confront details that shape the layers above that you do deal with. This is one reason why Crafting Interpreters is such a beloved book... LLMs are now at the core of tools we all use everyday... don't you want to know how they are made?" — Hacker News user libraryofbabel

While some critics question the practical utility of building models from scratch given the massive compute requirements of state-of-the-art systems, the consensus among students of the material is that the value lies in the conceptual grounding. Understanding RLVR and inference scaling allows practitioners to better interpret current trends in AI, such as why models may struggle with non-math domains or why certain training techniques lead to specific failure modes.

Resource Availability

The video course is available for free, with accompanying code repositories hosted on GitHub. For those seeking a deeper theoretical walkthrough, the material serves as a companion to Raschka’s book, Build a Reasoning Model (From Scratch), which includes additional exercises and detailed explanations. A condensed 90-minute conceptual overview is also available via LinkedIn Learning for those looking for a high-level introduction before committing to the full implementation series.

Sources

Related

  • Project
    rasbt/reasoning-from-scratch

    A code repository accompanying a book that teaches you to build a reasoning LLM from scratch, starting from a pre-trained model and adding reasoning capabilities via inference-time scaling, reinforcement learning, and distillation — all runnable on consumer hardware.

  • Project
    rasbt/LLMs-from-scratch

    Official code for the book *Build a Large Language Model (From Scratch)* – a step-by-step, from-scratch implementation of a GPT-like LLM in PyTorch, covering text preprocessing, attention, model implementation, pretraining, and finetuning. Runs on laptops, with optional GPU acceleration.

  • Dispatch
    LFM2.5-24B-A2B: Liquid AI's 24B MoE Reasoning Model for On-Device Agents

    Liquid AI's LFM2.5-24B-A2B is a 24B-parameter reasoning model with 2B active parameters that leads or matches larger MoEs on five of eight agentic benchmarks while decoding 184 tokens/s on an Apple M5 Max, though Hacker News commenters report mixed real-world coding reliability and licensing confusion.

  • Dispatch
    Prevent cognitive debt by manually retyping LLM-generated code – insights from Hacker News discussion

    Ankur Sethi prevents cognitive debt in personal projects by manually retyping LLM-generated code, a practice that Hacker News commenters debate as either a useful learning technique or an inefficient workaround.

  • Dispatch
    AI & Frontier Tech Roundup: Cheap Model Releases, Coding-Agent Memory, and the Physical-AI Data Moat

    This week's AI dispatch centers on MiniMax M3.1 Flash landing free across four platforms, DeepSeek open-sourcing its production inference stack, new memory and harness layers for coding agents, and robotics investors arguing that real-world data capture — not model capability — is physical AI's bottleneck.