MMOOZZGG.com

Explained

AI assistant

In short

A large language model (LLM) is a neural network trained on a huge amount of text that writes its answer one piece of a word at a time. The explainer covers the name, a definition in plain words and in law, how the model works and is trained, its parts such as parameters and tokens, and where it makes mistakes. It ends with three ways to start.

Key ideas

  • The transformer architecture, described on 12 June 2017, underlies almost all modern language models.
  • Parameters are the numbers inside a neural network that are tuned during training.
  • An LLM can write and shorten texts, answer from documents, translate and write code.
  • A glossary of 15 terms explains each one for a child, in plain words, for business and for a specialist.

Prepared by AI from the article

What is an LLM: a large language model and how it works

Easy12 min read11 sources

A large language model (LLM) is a neural network trained on a huge amount of text that writes its answer one piece of a word at a time. Almost all modern language models are built on the transformer architecture, described on 12 June 2017. Below: what the abbreviation stands for, a definition from law, how it works, how it is trained, the kinds of models and three ways to start.

  • IT
  • Large language models
On this page · 10 sections

What LLM stands for (and not to confuse with LL.M.)

LLM stands for Large Language Model. “Large” means there are billions of parameters inside: GPT-3 had 175 billion. “Language” means the model works with text: it reads a request and writes a continuation. So an LLM is a large language model.

In IT the term means both the model itself and the services built on it: chatbots, assistants, AI agents. When a contract or a job ad says “LLM integration”, it means connecting such a model to the company’s software.

Do not confuse it with LL.M. (Legum Magister, Master of Laws): this is a master’s degree in law, a lawyer’s second academic degree. The abbreviation is the same, so some searches for “llm” are really about legal education.

A definition in plain words and in law

In plain words: it is a program that guesses, from the start of a text, which piece of a word comes next. It guesses so well that the result is emails, summaries, code and answers to questions.

There is no database of facts inside the model. There are patterns of language learned from a huge amount of text, and that is why the model can be wrong in a confident tone.

Russian law has no term “large language model”, but it has two close ones. They come from Russia’s National Strategy for the Development of Artificial Intelligence, as amended by Presidential Decree No. 124 of 15 February 2024, paragraph 5.

Subparagraph “n”: “large generative models” can “interpret and create multimodal data (texts, images, video and the like)”. Multimodal means of different kinds. They must do this “at a level comparable to or exceeding the results of human intellectual activity”.

Subparagraph “o”: “large foundation models” contain at least 1 billion parameters and are used “to perform a large number of different tasks”. A modern language model fits both definitions: it creates text and serves as the base for many programs.

How an LLM writes an answer: in short

The model does not see words. First the text is cut into tokens: whole words, parts of words or symbols. Splitting rare words into parts (the BPE algorithm) was described by Sennrich et al. in 2015 for machine translation.

Each token turns into a set of numbers. Then the transformer layers with an attention mechanism decide which words of the request matter to each other: for example, what “he” refers to in a long sentence.

At the output, the model computes the probability of each possible next token, picks one, adds it to the text and repeats the step. A long answer is thousands of such steps, so it takes longer and costs more than a short one.

Hence a common question: why does the model answer the same request differently? At each step it picks a token with some randomness, and a small difference at the start changes all the text after it.

The “temperature” setting sets how boldly the model takes less likely tokens: the lower the value, the more predictable the text. All six steps on one diagram are covered on the page about how a neural network works.

What a model is made of: parameters, tokens, context

Parameters (weights) are the numbers inside a neural network that are tuned during training. Their number decides the hardware you need and the price of an answer: GPT-3 had 175 billion parameters in 2020, and the Russian AI strategy counts a model as a large foundation model from 1 billion.

Kaplan et al. showed in 2020 that the error of a language model falls by a power law as the number of parameters, the volume of data and the amount of compute grow. This work explains why developers kept growing the size of models and training sets for years.

Tokens are the unit of count. Cloud models charge for input and output tokens, and the length of a request and an answer is also measured in tokens, not words.

The context window is the amount of text in tokens that the model sees at one time: the instruction, documents, the conversation history and the answer itself. A large window does not mean an accurate one: in the work of Liu et al. (2023), models found the needed fact in the middle of a long text much worse than at the start or the end.

How LLMs are trained

Training goes in three stages. The first is pretraining: the model reads a huge corpus of text and learns to predict the next token. The result is a base model that can continue text but follows requests poorly.

The second stage is fine-tuning on examples: people write model answers to requests, and the model learns from these pairs. The third is learning from human ratings (RLHF, reinforcement learning from human feedback): people rank several answers of the model, and it is trained further on their choices.

Ouyang et al. showed the effect of the third stage in 2022. People preferred the answers of InstructGPT with 1.3 billion parameters to those of GPT-3 with 175 billion, though the first model is more than a hundred times smaller.

Companies almost never train a model from scratch. They usually take a ready one and give it their documents through search (RAG, an answer from found passages) or train it further on their own examples.

What kinds of LLMs there are

By access. Closed models work only through the developer’s service: in a chat or through an API. Open-weight models can be downloaded and run on your own server, so the data does not leave the company.

By purpose. A base model continues text, an instruct model follows requests, and a reasoning model writes a chain of intermediate steps before the answer and solves logic and calculation tasks more accurately.

By size. Small models fit on a laptop or one server and answer simple requests fast. Large ones are stronger on complex tasks but need expensive hardware or payment for tokens.

By type of data. Text models work only with text; multimodal ones also understand images, sound or video. The strategy’s definition of “large generative models” is about exactly these.

Models are compared by people’s votes and by tests. In the Arena ranking in 2026, the top US model scored 1503 points and the top Chinese model 1464 (Stanford AI Index 2026). How this ranking works is on the page about LLM Arena, and the places of the models are in the LLM ranking.

What an LLM can do and where it makes mistakes

It can: write and shorten texts, answer from documents, translate, sort emails and requests, write code, pull the needed fields from contracts. As early as 2020, GPT-3 solved new tasks from a few examples in the request, without extra training. Examples by department are on the page LLM for business.

In a company it looks like this: the sales team gets a draft answer to a client in a minute, a lawyer sees the disputed clauses of a contract, support answers at night from the knowledge base. The model makes the first version, and a person decides what goes out.

More and more often the model works inside an AI agent: a program that presses buttons and calls services itself. In the OSWorld test (an agent working at a computer), accuracy grew from about 5% in 2024 to 66.3% in 2025, against a human level of 72.35%.

Where it errs: the model can confidently state something false; this is called a hallucination. Kalai et al. (2025) explain the cause: in pretraining a false statement cannot be told apart from a fact statistically, and tests reward a guess more than an honest “I don’t know”.

Other limits: the model does not know events after its knowledge cutoff date and gets lost in long calculations. So in work with clients, a rule or a person checks the answer, and the model takes facts from your documents, not from memory.

How to start: a chat, an API or your own model

Chat. The easiest start is a chat service: ChatGPT opened to everyone on 30 November 2022, and since then there have been many such services. It fits personal tasks: an email, a summary, a plan, an idea.

API (application programming interface). The company’s program sends a request to the model and gets an answer, and you pay for tokens. This is how a model is built into a CRM, the 1C accounting system, a Telegram bot or a website; services with an API and payment methods are collected in the section about AI model APIs.

Your own model. An open-weight model is installed on your own server or computer: data does not leave the company and you do not pay for tokens, but you need hardware. Where to start on a laptop is covered in the review of local neural networks.

If the model will see client data, 152-FZ, Russia’s personal data law, applies: personal data cannot simply be sent to a cloud abroad; the law sets conditions for this. Your own model inside the company’s environment removes the issue; read more on the page about the law and data.

How to choose between the three paths: a chat fits for testing an idea in a day, an API for building the model into processes without your own hardware, your own model for work with sensitive data and a large flow of requests. Many companies combine paths: they give simple tasks to the cloud and process documents with personal data in-house.

For a company, the path starts not with choosing a model but with a task: where the model will save staff hours and money. First the tasks and the calculation of the effect, then the model and the way to launch it.

What you are looking for: five routes through the section

People look for language models with different goals: to understand a term, choose a model, run it on their own hardware, adopt it in a company or work with models as a profession. Below are five routes, each with three steps from the simple to real work.

Every page of the section rests on primary sources with a check date: from the 2017 transformer paper to the Stanford AI Index 2026 report. Links to them are collected at the bottom of the page in the block “Sources and dates”.

LLM glossary: 15 terms

In the site’s glossary, each term is explained on four levels: for a child, in plain words, for business and for a specialist, with an example and a primary source. Below are the main words from the category “Language models”: from token to transformer.

The definitions rest on the original works. The token is described in the glossary by the paper of Sennrich et al. on splitting words into parts, and the context window by the study of Liu et al. on losing facts in the middle of a text. All terms of the category are collected on a separate glossary page.

Questions and answers

Are an LLM and a neural network the same thing?

No. A neural network is a general class of programs that learn from examples: they recognize faces and speech and forecast demand. A large language model is one kind of neural network, trained on text and built on the transformer. How AI, machine learning and language models relate is covered on a separate page of the section.

Are an LLM and ChatGPT the same thing?

No. ChatGPT is a service of the company OpenAI, opened to everyone on 30 November 2022, and the company’s language models run inside it. The same model is often available both in a chat and through an API, and there are dozens of services based on different models.

How is an LLM different from an AI agent?

A model writes text in reply to a request. An AI agent is a program around the model: it gives the model tools (search, CRM, email), memory and a goal, and the model decides by itself which step to take next. An agent does a task in full, for example sorts a request and creates a deal, while the model inside handles understanding and text.

Can you run an LLM on your own computer?

Yes, open-weight models run on a laptop or a server through programs such as Ollama and LM Studio. The more parameters, the more video memory you need: small models run on an ordinary computer, large ones need a server with graphics cards. How much memory a specific model needs is worked out by the calculator in the section about your own AI server.

Which LLM is best?

It depends on the task; test on your own examples. A model strong in code may lose to another in Russian text or work with documents, and rankings show an average result on other people’s tasks. Take two or three dozen typical requests of your company and compare the answers of several models; the rankings described above give you a starting point.

Sources and fact-checking

Checked by the editors on 16 October 2026.

  1. arXiv, 12 June 2017 (arxiv.org)
  2. arXiv, 28 May 2020 (arxiv.org)
  3. arXiv, 23 January 2020 (arxiv.org)
  4. arXiv, 4 March 2022 (arxiv.org)
  5. Sam Altman, Reflections, 6 January 2025 (blog.samaltman.com)
  6. Decree of the President of the Russian Federation No. 124, 15 February 2024 (pravo.ppt.ru)
  7. Pravo.ru, 21 March 2018 (pravo.ru)
  8. arXiv, OpenAI, 4 September 2025 (arxiv.org)
  9. Stanford HAI, 13 April 2026 (hai.stanford.edu)
  10. arXiv, 31 August 2015 (arxiv.org)
  11. arXiv, 6 July 2023 (arxiv.org)

Found a mistake? Tell us

Publications

Contact us