Definition
A large language model is a high-capacity learned model trained on large token sequences to estimate or generate language. Common objectives include predicting the next token or reconstructing masked tokens. The resulting model can be adapted through prompting, examples in context, retrieval, fine-tuning, or additional training.
There is no universal parameter, data, or compute threshold at which a language model becomes "large." The label is relative to its technical era and usually signals enough scale and breadth to support several tasks rather than one fixed classifier. Size alone does not establish quality, reasoning, safety, or generality.
Language-model research predates transformers and modern LLMs. OpenAI's 2018 generative pretraining paper combined a left-to-right transformer language model with task-specific fine-tuning. The 2020 GPT-3 paper then demonstrated zero-shot and few-shot task behavior through instructions and examples in context. These were influential steps in the current LLM lineage, not the invention of all language modeling.
Distinguish it from nearby terms
- A language model estimates patterns in token sequences. An LLM is the large-scale, high-capacity subset.
- A chatbot is an interface and application around a model. It adds conversation state, instructions, moderation, and product behavior.
- An agent adds goal-directed control, tools, state, and action policies. The LLM may propose actions but does not own permissions by itself.
- A multimodal model processes more than one modality inside the model. Some LLM products are multimodal; others connect a text model to separate vision or audio tools.
Operational significance
Identify the exact model version, tokenizer, context limit, decoding settings, prompt, tool boundary, and provider behavior. Evaluate the task at the system level and the model under controlled conditions. Otherwise, model capability and application engineering become impossible to separate.
Check your understanding
Two assistants use the same LLM, but only one can search private files and send messages. Those permissions belong to the systems around the model, so their risk profiles differ even if the model weights match.