LLM Architecture Knowledge Base

A chronologically-organized reference of decoder-only LLM architecture innovations. Each entry: the math, a simple implementation sketch, tradeoffs, ablation results, and which disclosed production models adopt it — with citations.

This site catalogs meaningful architectural innovations in decoder-only LLMs from 2017 to today. Coverage stays at "medium deep-dive" depth: enough to understand the move and reimplement a sketch, not enough to be a paper substitute.

We follow a strict closed-model policy: no inferred architecture claims about undocumented models. Every adoption claim links to a paper, model card, or official report.

The 2026 consensus stack

One transformer block, six choices

The default frontier dense decoder of 2024–26 is a small, settled set of choices. Each labeled module below links to its entry.

One transformer block: pre-norm residual layout with attention (RoPE + GQA + optional QK-Norm) and FFN (SwiGLU), normalized by RMSNorm. x_ℓ x_ℓ₊₁ residual stream Pre-Norm RMSNorm Attention GQA + RoPE + QK-Norm optional (MLA replaces GQA in DeepSeek) + residual add Pre-Norm RMSNorm FFN SwiGLU ( DeepSeekMoE for sparse variants) + residual add repeat L times (typically 32–80)
The frontier dense-decoder block, ca. 2026. Residual stream flows down the left; attention and FFN branches read, transform, and add back. Each labeled box links to its entry.

The open frontier

Eight years of open-weight models

Every open-weight model with a disclosed parameter count tracked in this knowledge base, plotted by release date and total parameters. Color encodes the releasing organization; dense models are circles, mixture-of-experts models are stars.

dense MoE
1B10B100B1T10T2023202420252026release dateparameters (log)
40 open-weight models with disclosed parameter counts, plotted by release date and total parameters (log scale). Hover a point for details, click it to open the model entry, click a company chip to highlight that family.

Browse

By category