Large Language Model (LLM)

A large language model (LLM) is a deep neural network, almost always a transformer, trained on very large text datasets to predict the next unit of text given the preceding context. Through this objective the model learns grammar, facts, reasoning patterns, and stylistic conventions, which it can apply to tasks it was never explicitly programmed for.

How It Works

An LLM converts input text into tokens, maps each token to a vector, and processes the sequence through stacked transformer layers that use attention to weigh relationships between tokens. The final layer produces a probability distribution over the vocabulary for the next token. Generation repeats this step, appending each chosen token and feeding the result back in. Model scale is described in parameters, often billions, and capability tends to rise with more parameters, more data, and more compute.

Why It Matters

LLMs power chat assistants, code generation, summarization, classification, and retrieval systems. A single pretrained model can be adapted to many downstream uses through prompting, fine-tuning, or retrieval augmentation, which lowers the cost of building language features. Teams must still manage limitations such as hallucination, stale knowledge, context-window limits, and inference cost.

Related Terms

LLMs build on the transformer architecture and process text as tokens. They are often called foundation models when used as a base for many applications, and are adapted through fine-tuning and run during inference.