LLM (Large Language Model)

What does LLM mean?

Large: the model has a very large number of internal settings, called parameters, learned from a very large amount of text. Language: it works with human language (and often code). Model: it is a trained mathematical system, not a database of facts.

If you saw LLM in a text, a post, or a news story about AI, it almost certainly means large language model. The similar-looking LL.M. is a Master of Laws degree, which is unrelated.

How an LLM works

The analogy: autocomplete on your phone, scaled up enormously. Your keyboard guesses the next word from the last few; an LLM predicts the next piece of text from everything in front of it, over and over, until a full answer forms.

The precise version: text is split into tokens, small chunks such as words or pieces of words. Almost all modern LLMs use the transformer architecture, introduced in the 2017 paper "Attention Is All You Need," which lets the model weigh how every token relates to every other. Anthropic's documentation describes the first stage, pretraining, as training a model on a large body of text to predict the next word. Pretrained models are then fine-tuned, often with human feedback (RLHF), to become helpful assistants.

Training vs inference

Training is the expensive, one-time process of building the model. Inference is every time you use it afterward: you send a prompt, the model generates a reply token by token. Training sets what the model knows; inference is what you pay for per use and what determines speed, or latency.

Examples of LLMs

A concrete use: paste a long customer email into an LLM and ask for a three-sentence summary and a polite reply. It does both in seconds, from the same underlying ability to predict language.

• Claude, from Anthropic.
• GPT models, from OpenAI, which power ChatGPT.
• Gemini, from Google.
• Llama, from Meta, released as open weights that anyone can download and run.

Why LLMs matter

The LLM is the engine under almost every AI product you use: chatbots, writing tools, coding assistants, and AI agents, which wrap an LLM with tools and a loop so it can take actions. Knowing what an LLM is, and what it is not, is the foundation for using any of them well.

Related terms

• Token: the unit an LLM reads, writes, and bills in.
• Context window: how much an LLM can hold in mind at once.
• Foundation model: the broader category many LLMs belong to.
• Hallucination: confident wrong answers, a direct side effect of predicting plausible language.

Common misconceptions about LLMs

• An LLM looks facts up. On its own it generates text from learned patterns, which is why it can be confidently wrong. Search or RAG adds real lookup.
• An LLM remembers your past chats. The model itself has no memory between sessions; any memory is a feature the app adds around it.
• ChatGPT is the LLM. ChatGPT is an app; the LLM is the engine underneath it.
• Bigger is always better. Larger models often reason better, but smaller ones can be faster and cheaper and perfectly good for routine tasks.

LLM in practice

An LLM on its own talks; an agent built around one acts. See how that works in what an AI agent really is, and if you are picking one, our tutorial on how to choose an AI model covers the trade-offs.