LLM: definition, operation and examples (what is an LLM?)

This article was published on November 7, 2026, and is based on foundational research publications (Google, OpenAI), the CNIL glossary, the European Commission's AI Act pages, and publisher documentation consulted that day. Model names are those available at the beginning of November 2026.

An LLM, or Large Language Model, is a program trained on vast amounts of text to predict, word by word, the most likely continuation of a text. It's the engine behind ChatGPT, Claude, Gemini, Copilot, and Mistral's Vibe: when you ask a question, the assistant doesn't search for the answer in a database; it generates it by chaining together predictions. This mechanism explains both what LLMs do remarkably well (writing, summarizing, reformulating, and reasoning about a document) and what they do poorly (ensuring the accuracy of a fact). This article provides a simple definition, explains the four-step process, the vocabulary to know, the main models as of November 2026, and what an LLM concretely changes for a company.

The essentials in five lines

• An LLM is a large language model: a neural network trained on billions of texts to predict the next word in a sentence.
• He works by cutting the text into tokens (pieces of words), then calculating, token by token, the most probable sequence taking into account all that precedes it.
• Current LLMs they all rely on the Transformer architecture, published by Google researchers in 2017.
• The most commonly used in business In November 2026 are GPT-6 (OpenAI), Claude (Anthropic), Gemini (Google) and the Mistral AI models, often with open weights.
• Its main limitation is that he produces a plausible text, not a verified text: he can confidently invent a fact, a source, or a figure.

LLM: simple definition

The CNIL defines a language model as a "statistical model of the distribution of linguistic units" in a language, which can, for example, "predict the next word in a sequence of words." An LLM is simply a very large language model: it has been trained on such a large quantity of text (a large part of the public web, books, code, articles) that it has learned not only grammar, but also facts, styles, reasoning, and ways of arguing.

Three words of the acronym deserve to be taken literally. Wide : the model has tens or hundreds of billions of internal settings, called parameters. Language : it manipulates language, that is to say text, and by extension code and, for recent models, images, sound or video. Door Design It's a mathematical model, a function that transforms input text into output text. It has neither a database of answers nor access to the web by itself.

In French, one also encounters "grand modèle de langage," "grand modèle de langue," or "modèle de langage de grande taille." All refer to the same thing. We also speak of foundation model The CNIL (French Data Protection Authority) classifies under this name models "whose capabilities are general" and which can be adapted to a wide variety of tasks. An LLM (Lifecycle Management) model is the most common type of foundation model.

How an LLM works, in four steps

An LLM is not born an assistant. It becomes one in four stages, the first two of which take place at the publisher, before any release, and the last two each time you use it.

Stage	What is happening	What this changes for the user
1. Pre-training	The model reads trillions of text tokens and learns to predict the next token. This is the longest and most expensive step: weeks of computation on thousands of graphics processors.	The model "knows" how to write and knows many things, but only up to a cutoff date for its data.
2. Adjustment	It is then trained on examples of questions and correct answers, and then aligned with human preferences: evaluators compare several answers and the model learns to produce those they deem useful and safe.	The model responds to instructions instead of simply continuing a text, refuses certain requests, adopts an assistant's tone.
3. Review of your request	Your message, the tool's system instructions, the conversation history, and the attached documents are broken down into tokens and placed in the context window.	Anything that is not in this window does not exist for the model: hence the importance of providing the context.
4. Generation	The model calculates a probability for each possible token, chooses one, adds it to the text, then starts again, until the end of the answer.	The answer is displayed word for word. Two identical requests can give two different answers.

Step 2 is what made ChatGPT a success in November 2022. The technique, called reinforcement learning from human feedback (RLHF), was described by OpenAI in March 2022 in the article dedicated to InstructGPT: a smaller model but aligned with human preferences was preferred by the evaluators to a model one hundred times larger that was not.

A simple example illustrates step 4. To the sentence "Monday's meeting is postponed to," the model assigns a high probability to "Tuesday," "a later date," or "next week," and a near-zero probability to "giraffe." It hasn't grasped the concept of a meeting in the human sense; it has learned, from billions of examples, which outcomes are plausible. It is this capability, applied on a massive scale, that produces consistent responses across multiple pages.

Tokens, context, parameters: the vocabulary you need to know

Spas	Definition	Order of magnitude
Token	The unit of text that the model reads and produces: a short word, a piece of a word, a punctuation mark.	In English, OpenAI counts approximately 4 characters or three-quarters of a word per token; French uses a little more.
Context window	The amount of text the template can take into account at one time: your request, history, documents.	Up to a million tokens at Claude and Gemini, or several hundred pages.
Settings	The internal settings of the neural network, fixed during training.	175 billion for GPT-3 in 2020; publishers no longer publish this figure for their recent models.
Date of acquaintance	The date on which the training data stops.	Often several months before the model's release; web search partially compensates for this.
Inference	The calculation performed for each answer.	That's what APIs charge, per million input and output tokens.
Reasoning model	An LLM that first produces an intermediate reflection before answering.	Slower and more expensive, but more reliable on multi-stage problems.

These concepts have a direct practical consequence: a Language Learning Model (LLM) doesn't have memory in the human sense. It only remembers a conversation because the tool displays the history in the context window with each message. The "memory" functions of assistants are notes that the software stores and retrieves. This is why a longer conversation becomes more expensive and sometimes less accurate: the model has to reread everything that came before with each reply.

Where do LLMs come from: key dates

Date	Event	Why it matters
June 2017	Eight researchers, mostly from Google, publish "Attention Is All You Need", which describes the Transformer architecture.	The attention mechanism allows us to process an entire text in parallel and to link each word to all the others. All current LLMs are derived from it.
2018	OpenAI publishes GPT, Google publishes BERT.	Initial models pre-trained on large corpora and then adapted to specific tasks.
May 2020	OpenAI presents GPT-3, 175 billion parameters.	The model performs tasks based on a few examples given in the instructions, without retraining.
March 2022	OpenAI describes InstructGPT and alignment with human preferences.	The model follows instructions instead of continuing a text.
November 30, 2022	Launch of ChatGPT.	The general public discovers LLMs through a simple chat window.
2023	Meta publishes Llama, Mistral AI publishes its first open-weight model, Anthropic launches Claude.	The birth of a competitive market, with models that can be self-hosted.
September 2024	OpenAI launches o1, the first mainstream reasoning model.	The models take the time to "think" before responding.
2025 and 2026	LLMs become agents: they use tools, navigate, produce files and chain tasks together.	We no longer just converse with an LLM, we delegate work to them.

Examples of LLMs: the main models in November 2026

Each software publisher now offers a family of models, from the lightest to the most powerful, rather than a single model. The lightest responds in a second for a few cents; the most powerful takes longer to process for up to a hundred times the cost.

Editor	Models	Where they are used	feature
OpenAI (United States)	GPT-6 Luna, GPT-6 Sol, GPT-6.1 Sol, GPT-6 Astra	ChatGPT, Microsoft Copilot (in part), API	The broadest ecosystem, with agents, Codex and ChatGPT Work
Anthropic (United States)	Claude Haiku 5.5, Sonnet 5.5, Opus 5.5, Fable 5.1	Claude, Claude Code, CoworkCopilot (in part), API	Writing, long documents, coding and task delegation
Google (United States)	Gemini 3.6 Flash, Gemini 3.1 Pro, Gemini 4 Argon	Gemini, Google Workspace, Gemini Notebook, API	Integration with Gmail, Docs, and Meet, in a context of one million tokens
Mistral AI (France)	Mistral Small, Mistral Medium 3.5, Mistral Large	Vibe (formerly Le Chat), API, hosting on your servers	European publisher, open-weight models, on-site deployment possible
Open templates from other publishers	Llama (Meta), Qwen (Alibaba), DeepSeek	On-premises hosting, cloud platforms	Downloadable file sizes, to be assessed according to the license and origin

An "open weights" model is one whose parameters are published by the developer: it can be downloaded and run on your own servers, without the data leaving the company. This isn't always strictly open source, as the training data and complete code often remain closed, and the license may restrict commercial use. To choose the right weight range for your task, see our article The right AI model for each task.

LLM, generative AI, chatbot, agent: what are the differences?

These words are often used interchangeably. However, they refer to different, overlapping levels.

Spas	What it is	Example
Artificial intelligence	The entire domain: any system that performs a task associated with human intelligence.	A spam filter, a recommendation engine
Generative AI	Systems that produce new content: text, image, sound, video, code.	An image generator, an LLM
LLM	The engine of generative text AI.	GPT-6 Sol, Claude Opus 5.5, Gemini 4 Argon
Conversational assistant	The application that makes an LLM available, with an interface, instructions, tools and safeguards.	ChatGPT, Claude, Gemini, Copilot, Vibe
AI Agent	An LLM that takes action: it plans, uses tools, links steps together and delivers a result.	Copilot Cowork, Claudius Cowork, ChatGPT Work, a Copilot Studio agent

The distinction between the model and the application is more important than it seems. The same Claude model can be found in the Claude application and in Microsoft 365 Copilot, with different results, because the instructions, tools, and documents it has access to change. Everything surrounding the model has a name, the harness, which we explain in What is an AI harness?For the difference between generative AI and action-oriented AI, see Agentic AI: definition and examples.

What is the purpose of an LLM in business?

Understanding that an LLM predicts plausible text helps in knowing where to place it. It excels when form matters and content is provided; it is fragile when asked to provide the content itself.

What an LLM does well	He does it wrong without taking precautions.
Write a first draft based on notes	Cite a legal precedent, standard, or article of law from memory
Summarizing a long document or a meeting	Provide a recent figure without web search
Rephrase, change tone, translate	Calculate mentally on large spreadsheets (he must use a calculation tool)
Extracting information from a contract or report	To guarantee that a source he proposes exists
Compare several documents and list the discrepancies.	Knowing your internal processes if they are not provided to him/her
Write or explain code, an Excel formula	To make decisions for you on a high-stakes issue

The right-hand column is not a list of prohibitions: it's a list of cases where you need to provide the model with the correct documents, activate web search or a calculation tool, and verify. This is the purpose of the RAG technique, which makes the model respond using your own documents, detailed in RAG: definition, operation and limitationsAnd that's what we practice in our training. Mastering Generative AI, on the actual documents of the participants, regardless of the tool chosen by your company.

The limitations of an LLM to be aware of

• Hallucinations. The model can assert a false fact, invent a reference or a figure, with the same confident tone as if it were a true fact. This is a direct consequence of its operation: it produces the most plausible text, not the truest. We detail the causes and countermeasures in AI hallucinations: definition, causes, and how to reduce them.
• The date of acquaintance. Without web search, it doesn't know what happened after its training data stopped, and doesn't always report it.
• Confidentiality. Depending on the product and settings, what you paste into a consumer-grade assistant can be used to improve the models. Enterprise offerings commit to not training the models on your data. We detail the obligations in AI and GDPR.
• Biases. The model reflects the texts on which it was trained, with their inherent imbalances. For decisions affecting individuals (recruitment, evaluation), human oversight is essential.
• Cost and energy. Each response requires processing power. Instinctively using the largest template to rephrase an email is more expensive and consumes more resources, for the same result. See Frugal AI: definition and levers.
• Variability. Two identical requests do not always yield the same answer. For repeated use, stable instructions, examples, and proofreading are necessary.

LLM and AI Act: what concerns your business

The European regulation on AI refers to LLMs as "general purpose AI models". It distinguishes between two roles. providers These models—namely OpenAI, Anthropic, Google, and Mistral—have been subject to transparency and copyright compliance obligations since August 2, 2025: technical documentation, a summary of training data according to the model published by the Commission, and a copyright policy. The Commission targets models trained with more than 10²³ computational operations and capable of generating language; beyond 10²⁵The model is presumed to present a systemic risk, and its obligations are strengthened. Models placed on the market before August 2, 2025, have until August 2, 2027, to comply.

A company that uses ChatGPT, Claude, or Copilot is not a model provider. Its obligations depend on the use case: training its teams in AI since February 2, 2025 (Article 4), informing users when content or an exchange is generated by AI, and strict requirements only for high-risk uses, such as recruitment. Details, including the timeline, are in our article. AI Act (AI Act): What your company needs to do.

How to choose an LLM for your business

In most organizations, the question isn't "which LLM?" but "which assistant?" A model is rarely chosen in isolation; rather, a tool (Copilot, ChatGPT, Claude, Gemini, Vibe) is selected that provides access to a family of models. Five criteria guide the decision:

1. The existing ecosystem. A company using Microsoft 365 already has Copilot Chat; a company using Google Workspace already has Gemini. Starting from there avoids paying twice.
2. The data. Where are they processed, are they excluded from training, can the model be hosted in-house? For the most sensitive data, an internally hosted open-weight model may be justified.
3. Dominant practices. Writing and long documents, data analysis, coding, task delegation: the tools are not equal for every use.
4. The total cost. Licenses, usage-based consumption for agents and the API, but also training: an unused license is the most frequent cost.
5. Reversibility. The models change every three months. Well-written instructions, reference documents, and procedures can be transferred from one tool to another; dependence on a single publisher, less so.

To compare the assistants, see our comparisons. Claude or ChatGPT et Gemini or ChatGPTAnd to use them properly, our guide Prompt engineering: the method for writing effective prompts.

F.A.Q

What is an LLM, in one sentence?

An LLM, or large language model, is an artificial intelligence program trained on huge amounts of text to predict the next word, enabling it to write, summarize, translate, and answer questions in natural language.

What does LLM mean?

LLM is the English acronym for "Large Language Model," which is translated into French as grand modèle de langage or modèle de langage de grande taille. In a legal context, LL.M. also refers to a master's degree in law, unrelated to AI.

Is ChatGPT an LLM?

Not exactly. ChatGPT is an application, a conversational assistant. It operates using OpenAI LLMs, specifically the GPT-6 family (Luna, Sol, and Astra) as of November 2026. Similarly, Claude, Gemini, and Copilot are applications that rely on LLMs.

What is the difference between LLM and generative AI?

Generative AI refers to all systems that produce content: text, images, sound, and video. A Language Modeling (LLM) is the type of model that produces both text and code. All LLMs fall under the umbrella of generative AI, but an image generator is not an LLM.

Does an LLM understand what they write?

He doesn't understand in the human sense: he calculates probabilities based on sequences of words. But at this scale, the calculation reproduces reasoning, explanations, and analyses that are often relevant. In practice, he should be treated like a quick and brilliant collaborator who doesn't check his sources.

Is it possible to use an LLM without sending your data to the United States?

Yes, there are two ways: choose a publisher or service that processes data in Europe, or host an open-weight model, like those from Mistral AI, on your own servers or in a private cloud. The second option requires technical skills and computing infrastructure.

Sources

• CNIL, Glossary of artificial intelligence, entries "language model" and "foundation model", accessed November 7, 2026.
• Vaswani et al., Attention Is All You Need, June 2017.
• Brown et al., Language Models are Few-Shot Learners (GPT-3), May 2020.
• Ouyang et al., Training language models to follow instructions with human feedback (InstructGPT), March 2022.
• OpenAI Help Center, What are tokens and how to count them?
• Representation of the European Commission in France, EU rules relating to general-purpose AI models come into force, August 1, 2025.
• European Commission, guidelines on the obligations of providers of general-purpose AI models.