The structural anatomy of open-weight large language models · GPT-2 → 2026

The structural anatomy of open-weight LLMs

This atlas dissects the architectures of 80 open-weight large language models released between GPT-2 (2019) and 2026. Instead of benchmark scores, it records structure: how attention evolved from multi-head to grouped-query, multi-head latent and hybrid linear designs; how sparse mixtures-of-experts replaced dense feed-forward blocks; and which operators each laboratory adopted and when. Every figure is traceable to the public config.json shipped with the model or to the original publication.

80models in the corpus
10Family
55Sparse MoE
2019–2026Released

Nine ways into the same corpus — pick the view that answers your question.

Models A sortable index of 80 open-weight LLM architectures: parameters, active parameters, context window, attention variant, MoE layout, layers and heads. Evolution How LLM attention evolved from MHA to MQA, GQA, MLA and hybrid linear designs — and which operator each model adopted, with first adopters. Operators Every architectural operator in the corpus — attention, MoE, normalization, position and inference-side — with its mechanism, adoption count, first open-weight adopter and evidence. Timeline A chronological map of 80 open-weight LLM releases from 2019 to 2026, plotted by total parameters and colored by laboratory. Lineage Model families grouped by laboratory: OpenAI, Meta, Mistral, DeepSeek, Qwen, GLM, Kimi, MiniMax, StepFun and Hunyuan. Similarity Cosine similarity between model architecture feature vectors: attention class, MoE layout, normalization, position encoding, scale and context. Papers The primary publications behind every technique in the atlas: Transformer, GQA, MLA, RoPE, YaRN, RMSNorm, SwiGLU, Mixtral, DeepSeek-V2/V3 and more. Multimodal The vision-capable models in the corpus, identified objectively by vision fields shipped in their public config.json. Compare Compare up to four LLM architectures side by side: attention, MoE layout, layers, context, dtype and more. Map All 80 models on one map: release date versus total parameters (log scale), colored by laboratory. Wall Every model in the atlas as a card: family, date, parameters, attention class and MoE layout at a glance. Config Explorer Browse the full public config.json of any of the 80 open-weight models: every field, searchable and copyable. Quiz Test your knowledge of LLM architectures: attention variants, MoE layouts, laboratories and scale — questions generated from the atlas corpus.

Params · Active

1000B Kimi K3 Qwen3.8 2.4T A95B DeepSeek-V4-Pro Kimi K2 Base Kimi K2 0905 Kimi K2.5 Kimi K2.6 Hunyuan 4 Preview GLM-5.2 GLM-5.3 GLM-5 GLM-5.1 DeepSeek-V3 DeepSeek-V3 Base DeepSeek-R1 DeepSeek-V3.1 DeepSeek-V3.2 MiniMax-Text-01 Params Active
Largest open-weight models: grey bar is total parameters (log scale); red bar is the share active per token. · log₁₀