Examples
Worked examples
- Is an instance
GPT-4 (OpenAI)
- Is an instance
Claude 3 (Anthropic)
- Is an instance
Llama 3 (Meta)
Counter-examples
Looks similar, but isn't
- Not an instance
A BERT-based classifier fine-tuned only to label sentiment is not typically called an LLM in the generative sense
Editorial commentary
A large language model (LLM) is a neural network, typically built on the transformer architecture, trained on very large corpora of text (and, increasingly, other modalities) to predict and generate sequences of tokens. What makes a model “large” is not a fixed threshold but a combination of parameter count (typically billions to trillions), training-data scale, and compute budget large enough to produce broad in-context learning ability — the capacity to perform new tasks from a prompt and a few examples without task-specific retraining. This distinguishes LLMs from earlier task-specific NLP models (e.g. a single sentiment classifier), which are narrow by design.
For research-administration and disclosure purposes, the operational question is rarely “how large is large” — it is which specific instance of a model was used. The identifying attributes that matter for reproducibility and disclosure are: model family and version (e.g. GPT-4o, Claude Opus 5, Llama 3.1), the training-data cutoff date, whether the deployment used retrieval augmentation or fine-tuning on top of the base model, and the provider/deployment context (a vendor API vs. a self-hosted open-weights checkpoint). A generic statement that “an LLM was used” does not meet most journal or funder AI-disclosure requirements; a named model, version and date does.
Worked examples
- GPT-4o, Claude Opus 5, Gemini 2.5, and Llama 3.1 are all LLMs, differing in parameter count, training regime, licensing (open-weights vs. API-only) and modality support.
- A researcher disclosing AI use in a manuscript should name the specific model and version (“GPT-4o, OpenAI, accessed March 2026”) rather than the generic category.
What it is not
A narrow, single-task NLP model (a spam classifier, a named-entity recognizer trained on one dataset) is not an LLM even if it uses a transformer architecture, because it lacks the broad, general-purpose in-context capability that defines the category. See foundation model for the broader category LLMs sit within, and frontier model for the subset operating at the current capability/compute frontier.
References
- Bommasani et al., ‘On the Opportunities and Risks of Foundation Models’ (2021)
- Brown et al., ‘Language Models are Few-Shot Learners’ (NeurIPS 2020)
Also known as
LLM · Foundation language model
Machine-readable encodings
Use in your systems
<role vocab="credit"
vocab-identifier="https://casrai.org/dictionary/"
vocab-term="Large language model (LLM)"
vocab-term-identifier="https://casrai.org/dictionary/term/large-language-model" />{
"@context": "https://schema.org",
"@type": "DefinedTerm",
"@id": "https://casrai.org/dictionary/term/large-language-model",
"name": "Large language model (LLM)",
"identifier": "https://casrai.org/dictionary/term/large-language-model",
"description": "A neural-network model trained on large text corpora using self-supervised next-token prediction (or analogous objective), with parameter counts typically in the billions, capable of generating coherent text and performing a broad range of natural-language tasks without task-specific training.",
"inDefinedTermSet": "https://casrai.org/dictionary/domain/genai-disclosure#set",
"url": "https://casrai.org/dictionary/term/large-language-model",
"sameAs": [
"LLM",
"Foundation language model"
],
"license": "https://creativecommons.org/licenses/by/4.0/",
"publisher": {
"@id": "https://casrai.org/#organization"
},
"author": {
"@id": "https://casrai.org/#editorial-team"
},
"datePublished": "2026-05-21T01:54:31",
"dateModified": "2026-08-22T15:54:46",
"inLanguage": "en-GB",
"isAccessibleForFree": true
}
