LLM Explained adds an interactive atlas of model architectures

LLM Explained has added an Architecture Atlas to its free, open-source course of 43 interactive pages. The tool compares open-weight language models through shared architectural components, showing how choices around normalisation, positional encoding, attention and feed-forward layers distinguish families such as Llama, DeepSeek, Qwen and Gemma.

The atlas combines diagrams tracing transformer design changes, configuration-generated model schematics, a concept-by-model comparison matrix and a searchable gallery of 120 open-weight architectures. Its detailed comparisons use 14 model configurations to generate diagrams, parameter counts, KV-cache estimates and specification cards from the same underlying data. Models carry labels indicating checks against vendor configurations or model cards, or validation through reproduced parameter counts.

The comparison tools highlight differences between grouped-query and multi-head latent attention, dense networks and mixtures of experts, and local versus global attention. Gallery cards use colour-coded layer strips to distinguish attention patterns and hybrid designs. Users can search by model, company or attention type, filter architectures and inspect model specifications. Concept links connect unfamiliar components to course explanations, animations and runnable Python. Gallery data comes from a public snapshot of Sebastian Raschka’s LLM Architecture Gallery; the site generates its own diagrams rather than importing images or benchmark scores.