OPEN AI GLOSSARY
Back to glossary

LLM

Large language model

A large language model is a deep learning language model trained on very large textual datasets. It is a kind of Model; its complete application or agent runtime is a separate system. [8] [14] Editorial interpretation

On this page

Definition comparison

EU-U.S. terminology

Institutional; language models

The second edition emphasizes deep learning and extremely large text datasets, and includes models that generate text or form representations useful for downstream tasks. [8]

GPT-3 study

Peer-reviewed; a specific experimental scale

Brown and colleagues study an autoregressive model with 175 billion parameters alongside smaller models. Its training corpus mixes web sources, books and Wikipedia; that recipe is an example, not a definition. [9]

Scope

This page concerns learned language models, including text generation and representation uses. The word “large” needs a stated scale and comparison context. [8] [9] Editorial interpretation

Includes

  • A trained text-generation checkpoint is a model artifact; adding retrieval, tools and execution control creates a larger application around it. [13] [14] Editorial interpretation
  • Representation-producing language models are within the cited institutional usage; generation is not its only permitted behavior. [8]

Excludes and counterexamples

  • A corpus alone is training material, not a trained model. A tool-connected assistant is not identical to the language model inside it. [14] Editorial interpretation
  • No universal parameter cutoff is established by the sources used here. Report parameters, training tokens and the relevant comparison rather than inventing a threshold. [9] [14] Editorial interpretation

Different usages and open boundaries

  • The institutional definition foregrounds textual training. For a multimodal product, specify its language-model component and additional modalities instead of assuming the same definition covers the whole system. [8] [14] Editorial interpretation
  • Training a model and conditioning an inference call are different operations: the GPT-3 few-shot experiments add examples in context without gradient updates. [9]

Operational criteria

Repository test: identify the learned model artifact, language task and objective, available training-material evidence, and reported scale. Record unknown proprietary details as unknown. Separately identify any tokenizer, serving system, tools and controller. [14] Editorial interpretation

Concept relationships

Only is-a means a strict subtype. Uses and supports/orchestrates describe composition and control.

  • Model
    • LLM is-a

      Within learned AI-model usage: an LLM is a specialized language model, hence a Model. [8] Editorial interpretation

Workflow uses LLM

An LLM workflow uses an LLM for one or more tasks; many workflows use no language model. [13] Editorial interpretation

Harness uses LLM

In an LLM-based agent runtime; evaluation harnesses and other harness usages have different scopes. [4] [6] Editorial interpretation

Agent uses LLM

For LLM-based agents only; general agents need not use a language model. [11] [13] Editorial interpretation

Repository working definition

Repository · Proposed / not certified

Use LLM for a learned language-model component at a documented large scale. Avoid reducing the term to corpus size alone or using it as the name for a tool-using system. No minimum parameter count is ratified by the repository. [14]

No consensus decision recorded.

Sources and research gaps

Research notes

Scale gap: the institutional definition supplies qualitative scale, not a universal threshold; a concrete GPT-3 size cannot settle the category for all architectures.

  1. EU-U.S. Terminology and Taxonomy for Artificial Intelligence, Second Edition

    EU-U.S. Trade and Technology Council, Working Group 1. Tier 1 / Institutional terminology document

    Second edition, prepared for the sixth TTC ministerial in 2024. Printed page 12, Model and Large language model; page 8, Machine Learning. Accessed 2026-10-10.

    Context: AI-system terminology for transatlantic risk-management and standards cooperation.

    Verification: Original text and publication record checked 2026-10-10.

  2. Language Models are Few-Shot Learners

    Tom B. Brown and others. Tier 2 / Peer-reviewed conference paper

    Advances in Neural Information Processing Systems 33, 2020. Abstract; sections 2.1-2.2, model sizes and training datasets. Accessed 2026-10-10.

    Context: GPT-3 experiments and corpus; a concrete example, not a universal LLM definition.

    Verification: Original text and publication record checked 2026-10-10.

  3. Building effective agents

    Anthropic. Tier 3 / First-party engineering blog

    19 December 2024; live page includes subsequent tooling updates. What are agents?; Building blocks, workflows, and agents. Accessed 2026-10-10.

    Context: Company-specific distinction between predefined LLM workflows and dynamic agents.

    Verification: Original text and publication record checked 2026-10-10.

  4. Repository terminology guide

    Repository internal proposal. Tier 4 / Repository internal proposal; not certified

    Repository proposal 1.3 / not certified. Term-specific proposal sections and editorial policy. Accessed 2026-10-12.

    Context: Internal proposals and editorial interpretations, not external consensus.

    Verification: Versions 1.1 and 1.2 add the Schema proposal and the flow/governance experimental example while preserving prior proposal sections. All remain unratified; no adoption decision claimed. Version 1.3 changes branding and certification scope only; it does not certify any proposal.

Revision record

  1. 1.0 / 2026-10-10

    Researched and rewritten with attributed definitions, typed relations and an unratified Repository proposal.

  2. 1.1 / 2026-10-12

    Scope the working definition and certification labels to this repository under Open AI Glossary. Citations and proposed status remain unchanged.