What Is a Large Language Model?

A large language model, or LLM, is an AI model trained on enormous amounts of text to understand and generate language. It works by predicting the next piece of text, over and over, based on patterns it learned during training. LLMs power chat assistants such as ChatGPT, Claude and Gemini, and they can write, summarize, translate, answer questions and help with code.

What an LLM is

You may see it written as "an LLM" or "a LLM." Either way, LLM stands for large language model. IBM describes LLMs as "deep learning models trained on immense amounts of data, making them capable of understanding and generating natural language and other types of content."

Each part of the name tells you something:

• Large refers to scale: the model is trained on a vast amount of text and contains a very large number of internal settings, called parameters. OpenAI's GPT-3, released in 2020, had 175 billion parameters. Developers of today's most capable models generally no longer disclose parameter counts.
• Language refers to what it works with: words, sentences and other text, including computer code.
• Model refers to what it is: a trained mathematical system that turns an input into an output.

An LLM is a type of neural network, and most are built on an architecture called the transformer, introduced by Google researchers in 2017. The LATHIC Glossary has the short definition of a large language model.

How LLMs work

Text is broken into tokens

LLMs do not read text as whole words. They split it into smaller pieces called tokens, which can be whole words, parts of words or punctuation. Everything an LLM reads and writes is measured in tokens. What Is a Token in AI? explains why that matters for limits and pricing.

The model predicts what comes next

At its core, an LLM does one thing: given a sequence of text, it predicts which token is likely to come next.

Given "The capital of France is," a well-trained model assigns a high probability to "Paris." It adds that token, then predicts the next one, and the next, until it has produced a full response. OpenAI describes its models as using what they learned about how words typically appear together "to predict the next most likely word when generating a response, one word at a time."

Models do not always pick the single most likely token. A small amount of controlled randomness makes responses more natural and varied, which is also why the same question can get different answers.

Training builds the patterns

LLMs learn in stages:

1. Pretraining. The model processes a huge collection of text and repeatedly practices predicting the next token. OpenAI says the models behind ChatGPT learn from information publicly available on the internet, information accessed through partnerships with third parties, and information provided by users, human trainers and researchers. Along the way, the model's parameters come to capture patterns in grammar, facts, reasoning styles and writing formats.
2. Refinement. A model that only predicts text is not yet a helpful assistant. Developers train it further on examples of good responses and on people's ratings of its answers, so it learns to follow instructions. OpenAI's 2022 InstructGPT research is the classic example; today's methods add further kinds of training.

Attention helps it keep track of meaning

The transformer architecture uses a mechanism called attention, which lets the model weigh how strongly each part of the text relates to the other parts that came before it. That helps it work out, for example, what "it" refers to in a long sentence, or which earlier instruction a question depends on.

The context window sets a limit

An LLM can only consider a certain amount of text at once, including your messages, any documents you provide and its own replies. This limit is called the context window. Context windows vary widely between models, and information outside the window is not available to the model when it responds.

The model is not the same as the chatbot

An LLM is the model, and a chatbot like ChatGPT is a product built around it.

The same LLM can power many different products. Businesses often access models through an API, a way for software to connect to a service, to build their own tools. The product decides what the model is told, what tools it can use and what the user sees. The LATHIC Glossary describes the conversational version of a language model as a chat model.

What is the LLM in ChatGPT?

ChatGPT is OpenAI's chat application. The models underneath it are OpenAI's own, mainly from its GPT family. When OpenAI releases a new model, ChatGPT can switch to it, so the app you use stays the same while the model powering it changes. The app adds features on top, such as searching the web and working with uploaded files.

What LLMs do well

LLMs are strongest at tasks that involve understanding and producing language:

• drafting and editing writing in different styles
• summarizing long documents and conversations
• translating between languages
• answering questions and explaining concepts
• extracting key details from unstructured text, such as names, dates and action items
• writing, explaining and debugging code
• brainstorming ideas and working through problems step by step

LLM vs generative AI

Generative AI is the broad category of AI that creates new content, including text, images, audio and video. An LLM is one kind of generative AI, specialized in language. Nearly every LLM in use today is generative AI, while image generators and video generators are generative AI that are not LLMs. Many newer models blur the line by working with text, images and audio together. For the broader picture, see What Is Generative AI?.

LLMs vs AI in general

AI is the entire field of making computers perform tasks associated with human intelligence. An LLM is one specific type of AI model. Much of the AI people use, such as fraud detection, navigation and face recognition, does not involve LLMs at all. LLMs became the most visible form of AI because tools like ChatGPT let anyone interact with them directly in everyday language.

Where LLMs fall short

They can state false information confidently. OpenAI defines hallucinations as "plausible but false statements generated by language models." Its researchers argued in 2025 that one reason models hallucinate is that training and evaluation methods reward guessing over acknowledging uncertainty.

Their knowledge has a cutoff. An LLM's built-in knowledge comes from data collected before a certain date. Model developers publish these cutoff dates, and without a search tool, a model may not know about later events.

They do not check their own work. An LLM generates what is likely, not what is verified. Even when a product connects it to reliable sources, nothing in the process guarantees that an answer is true.

They can lose track in long tasks. Anything outside the context window is unavailable to the model, and performance can slip over long conversations or complex, multi-step work.

Their abilities are uneven. The International AI Safety Report 2026 notes that general-purpose AI can excel at complex tasks like generating code and answering expert-level questions, while struggling with tasks that seem simpler, such as counting objects in an image.

They reflect their training data. Biases and gaps in the text a model learned from can show up in its answers.

For why false answers happen and how to reduce them, see What Is an AI Hallucination?.

Why LLMs matter

LLMs are behind many of the AI tools people now use at work, from chat assistants to writing features in email, documents and customer service software. When a product says it is "powered by an LLM," a few questions clarify what that means:

• Which model does it use, and how recent is it?
• Can it access current information or your organization's documents, or is it working only from what it learned in training?
• What happens to the information you enter?
• How are its answers checked before anyone relies on them?