Large language model (LLM)

What it is

A large language model (LLM) is a computer program that has learned from enormous amounts of text how language is put together. Given a question or an instruction, it writes a fitting answer, one small piece of text at a time.

“Large” refers to two things: the amount of training text, often many trillions of words, and the number of adjustable settings inside the model, called parameters, which runs into the billions. GPT-3 from 2020 already had 175 billion parameters.²

How it works

Text becomes tokens. Text is cut into short pieces (tokens), roughly three quarters of a word each in English.
Training: predict the next token. The model reads text and is repeatedly asked to guess the next token. Each wrong guess slightly adjusts its parameters. After billions of rounds it has absorbed grammar, facts and styles.
The transformer architecture. Since 2017, almost all LLMs use the transformer, which lets every word in a text take every other word into account at once.¹
Fine-tuning for conversation. A raw model only continues text. To follow instructions, it is trained further on examples and on human ratings of good and bad answers (reinforcement learning from human feedback).³
Answering. When you ask something, the model generates its answer token by token, each one based on your question and everything it has written so far.

Short history

2017Google researchers publish the transformer architecture.¹
2018First large pre-trained language models: BERT (Google) and GPT (OpenAI).
2020GPT-3 shows that a single, very large model can solve many tasks from a few examples.²
2022OpenAI releases ChatGPT on 30 November; LLMs reach the general public.⁶
2023Open-weight models such as Meta’s Llama let companies run LLMs themselves.
2025EU obligations for providers of general-purpose AI models apply from 2 August.⁵

What businesses use it for

Drafting and summarising: emails, reports, offers, meeting notes.
Customer service: answering routine questions, with a human for the rest.
Searching own documents in plain language, usually by connecting the model to a company’s files (retrieval-augmented generation).⁴
Translation and localisation.
Writing and reviewing software code.
As the “brain” of AI agents that carry out tasks on their own.

Limits and risks

It can state false things convincingly. The model produces plausible text, not verified facts. This is called hallucination.⁷ Important outputs need a check.
Knowledge has a cut-off date. Without access to current sources, the model does not know recent events.
Data protection. What you type may leave the company. Which data may go into which service needs a rule.
Cost and speed. Price is charged per token; long documents and large models cost more.
Bias. The model reflects the patterns and gaps of its training text.

Rules in the EU

The EU AI Act treats LLMs as general-purpose AI models. Since 2 August 2025 their providers must, among other things, keep technical documentation, publish a summary of the training content and respect EU copyright; the most capable models carry additional safety duties.⁵ Companies that only use an LLM have their own, lighter duties, for example labelling AI-generated content.

Key terms

Token: A small piece of text the model reads and writes; prices and limits are counted in tokens.
Parameter: An adjustable number inside the model; together they store what it has learned.
Context window: How much text the model can take into account at once.
Prompt: The instruction or question you give the model.
Fine-tuning: Further training of an existing model for a specific task or style.
RAG: Retrieval-augmented generation: the model first looks up relevant documents, then answers from them.⁴
Hallucination: A confident but false or invented statement by the model.⁷

Latest research

New papers and reports that mention this term, found by our daily source scan. One line per source, quoted as published.