An LLM, or large language model, is an artificial intelligence (AI) model that understands and generates human language. It learns from vast amounts of data and answers a written instruction by predicting the most likely next piece of text, one token at a time. LLMs like these are the AI models behind ChatGPT, Copilot and Gemini, and they generate human-like text in seconds.
What you’ll take from this
In a few minutes you’ll know what an LLM is, how large language models work and how to use one with good judgment on your team. It is also where our AI certificate program starts.
- Understand what LLM stands for and how a language model differs from AI in general and from the app you chat with.
- Learn how LLMs are trained and why they sometimes get facts wrong with total confidence.
- Recognize the main types of large language models and the models behind the AI tools you already use.
- Apply a five-step method to put an LLM to work on a real task this week.
What does LLM stand for?
LLM stands for large language model. In simple terms, it is a type of AI model, built with deep learning techniques, that reads an instruction in natural language and writes a coherent answer. That makes LLMs the core of conversational assistants. In a different context, LL.M. is the abbreviation for a Master of Laws degree, but in AI it always means large language model.
What “large” means in a large language model
“Large” refers to two things. The first is the training data: books, websites, code and documentation, often in many languages.
The second is the number of parameters, the internal values the model learns and adjusts during training. Each parameter works like a small dial. Together they decide that after “the bank closed my” the likely next word is “account” and not “river”.
Scale grew fast. GPT-3, which OpenAI introduced in 2020, already had 175 billion parameters, according to its technical paper Language Models are Few-Shot Learners. LLMs today still have billions of parameters, and training LLMs at that scale takes enormous computational power. That scale lets them pick up tone, intent and context in a conversation.
How do LLMs work?
A large language model works in three phases. It is pre-trained to predict text, then fine-tuned to follow instructions, and when you use it, it generates the answer one token at a time. All of it runs on a neural network architecture called the transformer, which Google researchers introduced in the 2017 paper Attention Is All You Need.
Tokens: how a transformer neural network reads text
A language model does not read words the way you do. It splits text into tokens, which can be a word, part of a word or a punctuation mark.
Each token becomes a vector of numbers, similar to a word embedding. The attention mechanism then works out how each token relates to the rest of the sentence. That is how the model tells the “bank” you sit on from the “bank” that lends you money, a classic test of language understanding.
How are LLMs trained?
LLMs are trained using vast amounts of text and a single task: predict the next token. The text itself supplies the right answer, so no one has to label the data, a form of unsupervised learning. This is how models are trained before any tuning. Repeated billions of times, that task teaches the model grammar, vocabulary, common facts and the patterns of human language.
Nobody programs linguistic rules into it. LLMs are deep learning models, and their machine learning algorithms infer how language works from the training data, and that is what separates machine learning from traditional programming. Before training these models, data teams select and clean the dataset, because the quality of the data shapes the results. Pre-trained models can then be adapted to new tasks with far less data, a method known as transfer learning.
Fine-tuning: from model to assistant
A pre-trained model only completes text, because each parameter was tuned for prediction alone. To turn it into an assistant, providers fine-tune it on example conversations and then use reinforcement learning from human feedback (RLHF), where people rank the model’s answers. OpenAI described it in its 2022 paper on InstructGPT. After that step, it can answer questions, draft an email or summarize a contract.
Use: the answer is generated in real time
When you send a prompt, the model makes a prediction: it picks a likely token, adds it to the text and calculates again. That is why the same question can produce different answers, and why a clear prompt changes the result so much.
What are the different types of large language models?
Large language models, and the AI capabilities they offer, are usually grouped in four ways. The groups are base versus instruction-tuned models, open-weight versus proprietary models, text-only versus multimodal models, and standard versus reasoning models that work through a problem before answering. One model can belong to several groups at once.
| Type | What it means | Example |
|---|---|---|
| Base model | Pre-trained to complete text, not yet tuned to follow instructions | The first GPT models |
| Instruction-tuned model | Fine-tuned with example conversations and people’s ratings to act as an assistant | The models behind ChatGPT and Copilot |
| Open-weight model | The trained weights can be downloaded and adapted under the provider’s license | Meta’s Llama |
| Proprietary model | Reached only through the provider’s app or API | GPT, Gemini, Claude |
| Multimodal model | Works with images and audio as well as text | Google Gemini |
| Reasoning model | Spends extra steps working through a problem before it answers | The deep reasoning models available in Copilot’s Researcher agent |
Google, for example, says it built Gemini from the ground up to work across text, code, audio, image and video, according to its Gemini launch announcement (2023). Microsoft, in turn, lets Copilot’s Researcher agent run on deep reasoning models from OpenAI or Anthropic.
A short history of LLMs
The first LLM technology built on transformers appeared in 2018, a year after Google’s paper. In June, OpenAI published its first GPT model, which combined transformers with unsupervised pre-training, according to OpenAI’s announcement. In October, Google released BERT, a model built to understand text.
The breakthrough with the general public came on November 30, 2022, when OpenAI introduced ChatGPT. Since then, advanced LLMs have learned to work with images and audio and moved into everyday work tools such as Word, Excel and search engines.
What is the difference between GPT and an LLM?
GPT is a family of LLMs made by OpenAI, so every GPT model is a large language model, but not all LLMs are GPT models. ChatGPT, in turn, is the app built on top of OpenAI’s GPT models. So when people ask what LLM stands for in ChatGPT, the LLM is the GPT model inside the app.
The model is the engine. The app is the product you talk to, and it adds the interface, memory, web search and other tools depending on the plan. The same split applies to Gemini, Google’s assistant, which runs on models developed by Google DeepMind.
Microsoft 365 Copilot is powered by models from OpenAI. Since September 2025 it can also use Anthropic’s Claude models, starting with the Researcher agent and Copilot Studio, according to Microsoft’s announcement. Choosing between the two most common assistants? Compare Copilot vs ChatGPT with real use cases.
What are some examples of LLMs?
The most widely used LLM providers today are OpenAI with its GPT models, Google with Gemini, Anthropic with Claude and Meta with Llama. The first three are proprietary models you reach through an app or an API.
Meta releases Llama with open weights, so other companies can download and adapt it. It does so under its own license, with usage conditions set by Meta that are worth reading before you build Llama into a product.
| Model family | LLM provider | Where you find it | Access |
|---|---|---|---|
| GPT | OpenAI | OpenAI’s app and Copilot | Proprietary, via app or API |
| Gemini | Gemini app and Google Workspace | Proprietary, via app or API | |
| Claude | Anthropic | Claude app and Copilot | Proprietary, via app or API |
| Llama | Meta | Built into other companies’ products | Open weights, own license |
What are LLMs used for? Use cases at work
Teams use LLMs to generate drafts and handle other language tasks. The most common LLM applications at work fall into a few groups, one for each use case: text generation, summarizing, language translation, classifying documents and customer service. Developers can also integrate LLMs into internal systems through an API so they work with company data. AI agents then chain several of these steps together.
| Task | What the model does | Example on a team |
|---|---|---|
| Writing | Generates text from notes | Sales proposals, client emails, product descriptions |
| Summarizing | Condenses long documents into key points | Meeting notes, contracts, 40-page reports |
| Customer service | Answers questions in natural language | Virtual assistants that answer outside business hours |
| Text analysis | Sorts large volumes of text by topic or sentiment | Surveys, reviews, support tickets |
| Translation | Translates and adapts tone across language barriers | Emails to international clients, documentation |
| Code | Writes and explains formulas or scripts | Excel formulas, SQL queries, macros |
| Multi-step work | Chains steps and uses tools | AI agents that build a report from several sources |
How many companies already use AI
Official statistics measure AI in general, not LLMs alone. Even so, the most common use matches what large language models do best. According to Eurostat, 20.0% of EU enterprises with 10 or more employees used AI in 2025, and the most common use was analyzing written language (11.8%).
Adopting AI is easier than getting value from it. In McKinsey’s The state of AI in 2026, nearly nine in ten respondents report regular use of AI in at least one business function. Only 44% say AI is scaling across their enterprise. Closing that gap depends on how models use your instructions: people need to know what to ask and how to check what comes back.
What are the advantages of LLMs and their limits?
These models offer one main advantage, versatility: one model drafts, summarizes, translates or analyzes just by changing the instruction. You can also adjust tone and format without writing code. Their limits are well known. The risk grows when someone hands over complex tasks and nobody checks the output, so treat every answer as a first draft.
- Hallucination: because they generate the most likely text, they sometimes produce false figures, quotes or names with total confidence. Always check the facts you plan to use.
- Knowledge cutoff: they only know what was in their training data, unless they are connected to search or to your documents.
- Privacy: do not paste client data or confidential information into tools your company has not approved.
- Bias: they can repeat stereotypes present in the text they were trained on.
In the European Union, providers of these models also have transparency obligations under the AI Act. We explain them in what is a GPAI model.
Are LLMs actually AI? Where NLP fits
Yes. AI is the whole field, and generative AI is the branch of AI that creates new content such as text, images, audio or video. Language models are the generative AI systems that specialize in text. Most people meet them first through business applications such as email, documents and chat. An image generator is generative AI too, yet it is not a language model, because it outputs pictures rather than text.
Within natural language processing (NLP), the discipline focused on how machines process and understand language, these models changed the approach. Each task, from translating to classifying or summarizing, used to need its own model. Now one foundation model covers many of them, although specialized models still perform better on very narrow tasks.
How to use LLMs at work: a five-step prompt method
Use these five steps to test a first workflow that fits your business needs. You can have it running within a week.
- Pick a writing task you repeat every week, such as summarizing meetings or answering similar requests.
- Write the prompt with four parts: role, goal, context and output format.
- Review the result line by line the first time and note what you had to fix.
- Update the prompt with those fixes and save it for your team.
- Use the tool your company has approved, such as Copilot, whenever you work with internal data.
For example, a sales team that summarizes every call with a fixed prompt moves from writing notes to reviewing them. Here is a prompt you can copy and adapt.
You are a sales assistant. Summarize this transcript in five points: customer need, budget, objections, next step and agreed date. If any item is missing, write "not mentioned". Transcript: [paste the text here]
Before you save the summary to your CRM, check the date and budget against the transcript, because those are the two details worth verifying every time.
Frequently asked questions
These are the questions that come up most often when people start working with language models.
What is an LLM in simple words?
LLMs are AI systems that have read a huge amount of text and learned to predict which words come next. That lets them understand your question and write an answer, which is why they power assistants such as Copilot and Gemini.
Is ChatGPT an LLM or generative AI?
Both, in a sense. ChatGPT is a generative AI app, and the models it uses are LLMs from OpenAI’s GPT family. The app adds the interface, memory and tools around the model you talk to.
What is the difference between an LLM and a chatbot?
The LLM is the model that generates the text, and the app you type into is the product built around it. Most of these apps run on an LLM, but they add rules, memory and connections to other sources of information.
Do I need to code to use an LLM?
No. Clear instructions in plain English are enough. You only need to code if you want to build the model into your own applications or automate processes with company data.
Can LLMs make mistakes?
Yes. They generate the most likely answer, not a verified one. That means they can invent facts with great confidence, so check figures, names and quotes before you use them and ask the model to say where each fact comes from.
Which LLM does Copilot use?
Microsoft 365 Copilot runs on OpenAI’s models and also lets organizations use Anthropic’s Claude models in some features, such as the Researcher agent and Copilot Studio. Which model answers depends on the feature and on your organization’s settings.
Further reading
To follow how LLMs and their regulation evolve, start with these three sources.
- Stanford HAI AI Index Report, the annual reference report on the state of the field.
- Fundamentals of generative AI, a free training module on the Learn platform.
- The EU regulatory framework for AI, with the obligations of the AI Act.
Knowing what an LLM is comes first, and applying it to your work with good judgment comes next. In the Founderz Certificate Program in Artificial Intelligence and Innovation, developed in collaboration with Microsoft, you work with ChatGPT, Microsoft 365 Copilot and Copilot Studio. You also learn to build AI automations, workflows and agents. You finish with the Microsoft + Founderz Certification and earn official Microsoft badges along the way.
