Sebastian Raschka Details Manual Workflow for Analyzing Open-Weight LLM Architectures

Sebastian Raschka has documented his manual workflow for understanding large language model architectures, focusing specifically on open-weight models. His process begins with official technical reports, which he notes have become less detailed for many industry lab models. When weights are available on the Hugging Face Model Hub and supported by the Python transformers library, Raschka inspects configuration files and reference implementations directly to uncover architectural details. He emphasizes that this approach doesn't apply to proprietary models like ChatGPT, Claude, or Gemini. The workflow is intentionally manual rather than automated, as Raschka believes hands-on examination remains one of the best exercises for learning how these architectures function. He developed this methodology to create the LLM architecture sketches and drawings featured in his articles, talks, and the LLM-Gallery.