LLM Anatomy Atlas The structural anatomy of open-weight large language models

About this atlas

The LLM Anatomy Atlas is an independent, non-commercial reference that documents the structural design of open-weight language models. It deliberately records architecture rather than benchmark scores: attention variants, sparsity, normalization, positional encoding and provenance.

Methodology

Every structural figure is taken from the config.json file distributed with the model's open-weight release on Hugging Face, or — where no weights were published — from the architecture tables of the original paper. Derived classifications (MHA/GQA/MLA/linear-hybrid, dense/MoE) follow the field's standard definitions. Nothing is estimated: if a value could not be verified, it is left blank.

Sources

Primary sources are the public model repositories on Hugging Face (config.json), the arXiv papers listed on the Papers page, and the models' technical reports. Attention, RoPE, RMSNorm, SwiGLU and MoE techniques cite their original publications.

Data freshness

The corpus is re-synced against upstream releases on a regular schedule; the snapshot date is shown in the footer of every page.

Scope and disclaimer

This is an educational reference maintained by independent researchers and is not affiliated with, or endorsed by, any model laboratory. Model names remain the trademarks of their respective owners. Data reflects the snapshot date shown in the footer; newer revisions may differ.