LLM

An LLM (large language model) is a trained machine-learning model that maps text to text by predicting sequences of tokens. When extended to accept or produce additional modalities (images, audio, video), the result is a multimodal model; the core LLM mechanism remains token prediction over text.

Details

This repository assumes LLMs follow the transformer architecture and are trained using next-token prediction (see tokenizer) on large datasets. Most deployed LLMs are instruction following variants created via post-training such as fine-tuning. Compact variants are called small language models (SLMs). LLMs may be proprietary (accessible only through an inference provider API) or released as open-weight models that can be downloaded and run independently.

An LLM processes inputs within a bounded context window (see context size); the information an application expects the LLM to use is limited to what is placed into its context via the prompt (instructions, history, retrieved content, tool outputs). Across tasks, outcomes often depend on both knowledge and reasoning.

Synonyms

large language model