What is an LLM? Large language models explained

An LLM — a Large Language Model — is an AI trained on enormous amounts of text to predict which word comes next. It is the technology behind ChatGPT, Claude and Gemini, and the reason they can write, summarise, translate and reason in a way that feels human.

Below we cover what the abbreviation means, how a language model actually works, which models matter in 2026, and where the limits are. If you want the bigger picture first, see what AI is and how artificial intelligence works.

What does LLM stand for?

LLM stands for Large Language Model. Every word in the abbreviation carries weight:

• Large refers to two things at once — the volume of text the model was trained on, and the number of parameters it has learned. Both are measured in the hundreds of billions.
• Language — the model works with language. It reads text and produces text.
• Model — a statistical model, not a database. It does not look answers up; it computes them.

ChatGPT is therefore not an LLM in itself. ChatGPT is the product; the LLM is the engine under the hood. The same goes for Claude and Gemini — the interface is one thing, the model behind it another.

What is a language model?

A language model is a program that has learned how likely one word is to follow another. The idea is not new — the autocomplete on your phone's keyboard is a simple language model, and research in the field dates back to the 1950s. What is new is the scale.

A traditional language model looked a few words back and guessed the next one. A large language model holds tens of thousands of words in mind at once, has learned from a substantial share of everything written on the internet, and has enough learned connections to hold a thread through an entire document.

Strictly speaking, LLM is a subset: every LLM is a language model, but not every language model is large.

How does a language model work?

Here is a simplified explanation of how it works.

The model is trained by reading an enormous amount of text — books, articles, websites, code. During that process it learns language patterns, word usage, grammar, and even style and tone.

When the model then receives a question, it analyses not just the individual words but also the context. It looks at the words' relationship to each other to understand what the sentence actually means.

Based on what it has learned and the prompt you wrote, it generates an answer — one word at a time, where each word is chosen according to what is most likely to follow the previous one.

In other words: a language model is a very advanced version of the auto-suggestions on your phone. But it lacks consciousness and real understanding, and its knowledge is limited to the patterns it was trained on.

The model does not learn in real time from new conversations. It can be updated through new training, and it can be given access to fresh information via web search or your own documents.

Token, parameter and context window — three words you need

Almost every discussion of language models uses three terms. They are simpler than they sound:

• Token: the model does not read letters or whole words, but tokens — chunks of text roughly 3–4 characters long. "Language" might be two tokens, "hi" one. An A4 page of English text is roughly 500–700 tokens. When you use a model via an API, the price is almost always counted per token.
• Parameter: the learned weights in the model — what constitutes its "knowledge". The largest models have hundreds of billions of parameters. More parameters is not automatically better; the quality of the training data matters at least as much.
• Context window: how much text the model can hold in mind at once, measured in tokens. Everything you have written in the conversation plus the model's replies fits in here. When the window fills up, the earliest content starts falling out — which is why long chats can "forget" what you agreed at the start.

From raw data to finished chatbot: how an LLM is trained

A useful language model is built in three steps, and it is only in the last steps that it becomes something you want to talk to:

1. Pre-training: the model is fed enormous amounts of text and learns to predict the next token. This is the expensive, time-consuming part — it costs hundreds of millions of kronor in compute.
2. Fine-tuning: the model is trained further on more specific examples, such as dialogue, so it becomes good at actually answering questions rather than just continuing a text.
3. Human feedback (RLHF): people rate which answers are best, and the model is adjusted accordingly. This is the step where it learns tone, helpfulness and where the boundaries lie.

On top of this comes what makes today's models practically useful: tools, web search and the ability to connect your own documents.

Examples of language models in 2026

Most of the major model families build on the same underlying idea, but differ in size, price and what they are best at:

• The GPT series from OpenAI — the models behind ChatGPT. The latest generation comes in three tiers with different capacity and pricing; we go through them in our article on GPT-5.6: Sol, Terra and Luna.
• Claude from Anthropic — known for long context windows, careful analysis and strong coding ability.
• Gemini from Google — deeply integrated into Google Workspace and strong on multimodality.
• Llama from Meta and Mistral from France — open models whose weights you can download and run yourself, which makes them interesting when data is not allowed to leave the building. We have listed the best open models of 2026.
• DeepSeek and Qwen — Chinese models that have shown you can get close to the top for a fraction of the training cost.

The differences in practice are often smaller than the marketing suggests.

LLM, chatbot or AI agent — what is the difference?

The terms get mixed up constantly, but they sit at different levels:

• LLM is the model itself — the engine that predicts text.
• Chatbot is the interface you talk to, built on top of one or more LLMs.
• AI agent is an LLM that has been given tools and a goal, and can carry out several steps on its own: search, read, write, update a system. We cover what an AI agent is and how it works separately.

This is why the question "which one is best" is rarely answerable at the model level — the experience is decided just as much by the product around it.

What are the limitations of an LLM?

A language model lacks consciousness and real understanding. It computes probabilities; it does not actually "know" anything. The limitations worth keeping in mind day to day:

• Hallucinations: the model can invent sources, figures and quotes that look entirely credible. Always check facts that matter. There is now another kind of model that cannot do this — we explain it in what Jev is, the AI model that cannot make things up.
• Knowledge cut-off: the training data has an end date. Without web search, the model knows nothing about what happened after it.
• Bias: the model reflects the patterns in its training data, including the skewed ones.
• Confidentiality: what you paste in may be used for training in some services. Read the terms before sensitive material goes into a chat.
• Maths and arithmetic: language models are built for language, not arithmetic. Without tools, answers to exact calculations are unreliable.

The difference in 2026 is that most major models can now search the web in real time, reason step by step before answering, and handle images, audio and video — not just text. If you need fresh facts, you can simply ask the model to search, or use a tool like Perplexity.

How language models have evolved

Since this article was first written, language models have taken several big leaps:

• Larger context windows: models can hold far more text in memory at once — entire documents, codebases, or long conversations.
• Reasoning models: a new generation "thinks" problems through step by step, giving better answers on maths, code and logic.
• Multimodality: text, image, audio and video in the same model.
• Agents: models are connected to tools and can carry out tasks, not just answer.

The underlying principle is still the same, though: the model predicts the next word — just with a lot more power behind it now.

If you want a model that answers from your own documents rather than from everything it read on the internet, that works without code — we cover how to create your own GPT in 2026 and what applies now that custom GPTs are being retired.