What is an LLM? Large language models explained

An LLM, short for large language model, is a program trained on vast amounts of text to predict, piece by piece, the most likely continuation of a text. It is the engine behind ChatGPT, Claude, Gemini and Le Chat, and understanding how it works helps you use it better and spot its mistakes.

An LLM in one simple picture

Think of the predictive text on your phone. You type “Running late, be there in”, and it suggests “five” or “ten”. An LLM does the same thing with far more finesse: it takes into account everything that came before, sometimes hundreds of pages, and it learned from more text than any person could read in a lifetime.

Each letter has a meaning. Large refers to its size, Language to what it handles, and Model to what it is: in artificial intelligence, a model is a set of calculations whose settings were tuned from examples. If you are starting from scratch, begin with our guide to what artificial intelligence is, which places LLMs within the wider AI landscape.

How an LLM builds an answer

When you send a question, the model always follows the same five-step routine:

1. Split your text into tokens, small chunks that match a word, part of a word or a punctuation mark.
2. Convert each token into numbers, the only thing a computer can work with.
3. Score every possible next token with a probability, weighing how much each earlier word matters. This is the job of the attention mechanism, at the heart of the Transformer architecture.
4. Pick a token, with a small, adjustable dose of randomness. That is why the same question does not always get exactly the same answer.
5. Repeat with the new token added to the text, until the answer is complete.

Take “The cat is sleeping on the…”. The model rates “sofa”, “bed” or “windowsill” as very likely and “fridge” as much less so. It does not look the answer up in any database of facts: it calculates what sounds plausible. That one idea explains both its talents and its flaws. To follow the same process from the user’s side, read how ChatGPT works, explained simply.

Check yourself Does an LLM look up a database of facts to answer you?

No. The model on its own produces the most plausible text based on what it retained during training. To rely on verifiable facts, the assistant around it has to feed it documents or web search results.

How an LLM is built

Building a large language model happens in three broad stages, which differ hugely in time and cost.

Stage	What happens	What the model gains
Pre-training	The model reads an enormous volume of text (web pages, books, code) and practises guessing the next word, billions of times	Language, general knowledge, writing styles, ways of reasoning
Fine-tuning	It trains on sample dialogues, then people compare its answers and flag the best ones	Following instructions, an assistant’s tone, refusing some dangerous requests
Deployment	The now-frozen model answers users’ requests	Nothing new: each message triggers a computation that takes time and uses energy

Pre-training is the heaviest stage: it is where the model picks up most of what it knows. Fine-tuning, notably with RLHF (reinforcement learning from human feedback), then turns a “raw” model that can continue any text into an assistant that answers a question politely.

The name GPT sums up this story: Generative Pre-trained Transformer. The Transformer architecture was introduced by Google researchers in the paper “Attention Is All You Need”, posted on arXiv on 12 June 2017 (accessed 22 October 2026). Almost every current LLM descends from it.

Why “large”?

The word covers two things: the amount of text read during training, and the number of parameters in the model. Parameters are the internal numbers adjusted during learning, like billions of tiny tuning knobs. What a model “knows” is stored there, as numbers, not as a copy of the texts it read.

More parameters does not automatically mean better. A smaller model can be enough for a simple task, answer faster and cost less. Size is mostly useful for estimating how much memory you need to run the model.

Which LLMs are around today?

Dozens of LLMs are available. They differ by developer, price, how much text they can handle at once, and one key point: are their weights published or not?

Model	Developer	Release	Context (tokens)	Input ($/M tokens)	Output ($/M tokens)	Open weights
Claude Opus 5.5	Anthropic	September 22, 2026	1 M	$4	$20	No
GPT-6 Astra	OpenAI	September 4, 2026	1.05 M	$10	$50	No
Gemini 4 Argon	Google DeepMind	October 13, 2026	1.05 M	$2	$10	No
Mistral Large 4	Mistral AI	October 6, 2026	1.05 M	$1.36	$4.18	No
Kimi K3	Moonshot AI	July 16, 2026	1.05 M	$3	$15	Yes
DeepSeek V4.1 Flash 1017	DeepSeek	October 17, 2026	1.05 M	$0.15	$0.60	Yes

Two families stand out:

• Closed models, such as those from Anthropic, OpenAI or Google: you can only use them through the developer’s own app or API.
• Open-weights models, whose files can be downloaded. According to models.dev as of 22 October 2026, that is the case for Kimi K3, built by Moonshot AI. You can then run it on your own machines, within the limits of its licence.

Another criterion shown in the table is the context window: the maximum amount of text the model can take into account at once, measured in tokens. According to models.dev as of 22 October 2026, Claude Opus 5.5 accepts one million tokens. To dig deeper, the model comparison tool sets prices, context and public results side by side.

LLM, assistant, agent: three different things

These words often get mixed up, yet they describe three distinct layers:

• The LLM is the engine. Text goes in, text comes out.
• The assistant (ChatGPT, Claude, Gemini…) is the whole car: an interface, a history, a memory, web search, file reading and safety rules, built around one or more LLMs.
• The agent goes further still: it uses the LLM to plan, then chains actions with tools (browser, email, spreadsheet) until it reaches a goal.

In practice, two assistants running on the same model can behave differently, depending on what has been built around it.

A concrete example: you ask “What time does my local library open this Saturday?”. The LLM on its own has no idea and may invent plausible opening hours. An assistant with web search will first check the library’s website, then hand the page to the model so it can extract the answer. An agent could go one step further and book a study room online, if you allow it to.

What LLMs do well, and where they fall short

Large language models shine at anything to do with the shape of text: drafting an email, summarising a report, translating, rephrasing for a ten-year-old, brainstorming, explaining a concept, writing or fixing code.

Their limits follow directly from how they work:

• They can make things up. Because they produce what is plausible rather than what is verified, they sometimes state false things with confidence. These are known as AI hallucinations.
• Their knowledge has a cut-off date. Without web search, a model knows nothing about what happened after its training ended.
• Their working memory is capped by the context window: beyond it, part of your documents is ignored.
• Exact arithmetic is not their strength, unless they can call a calculator tool.

Going further

LLM jargon (tokens, context, temperature, inference…) keeps popping up in the news. Our glossary of the 10 AI words you need to know explains them one by one. And to see what a piece of text looks like in tokens, these site tools run the numbers right in your browser: