# Transformer

> The neural-network architecture, built on self-attention, that underlies virtually all modern large language models.

_The Agentic Web Lexicon · /glossary/transformer · [JSON](/api/glossary/transformer) · [all The Agentic Web Lexicon](/glossary)_

- **term:** Transformer
- **category:** knowledge-memory
- **short_def:** The neural-network architecture, built on self-attention, that underlies virtually all modern large language models.
- **long_def:** The transformer processes a sequence in parallel using self-attention, letting every token weigh its relationship to every other token. Introduced in 2017, it replaced recurrent architectures and made today's large language models — and the agents built on them — possible.
- **see_also:** context-window, fine-tuning, embeddings
- **etymology_origin:** — (verify-against-primary-at-build)
- **related_to:** context-window, fine-tuning, embeddings, tokenization
- **contrast_with:** Unlike earlier recurrent networks that read tokens one at a time, the transformer attends to the whole sequence in parallel, which is what enabled scaling to large context windows.
- **example:** Every frontier model in the Model Matrix — Claude, GPT, Gemini — is a transformer at its core.
- **source:** https://en.wikipedia.org/wiki/Transformer_(deep_learning_architecture)
- **status:** active
- **why_it_matters:** The transformer is the engine under every agent on the agentic web; its self-attention is why context — and how cheaply your content fits in it — matters so much.
- **sameAs:** https://en.wikipedia.org/wiki/Transformer_(deep_learning_architecture)
- **bridge_entity:** context-window
- **last_verified:** 2026-07-06
- **md_twin:** /glossary/transformer.md