What exactly is an LLM?
A Large Language Model, abbreviated LLM, is a computer program that understands language and generates language. When you chat with ChatGPT, revise an email with Claude, or ask Gemini a question, you're talking to an LLM. The term sounds clunky, but it describes pretty precisely what these models are: large language models.
The interesting word in there is large. Because "large" here doesn't mean "lots of storage" or "lots of features," but a very specific thing we'll unfold in a moment.
Imagine a person who has done nothing but read their entire life. Really nothing else. Wikipedia, novels, forums, textbooks, recipes, court rulings, Twitter, Reddit, programming code. At some point, this person develops an extremely good feel for which word comes next, no matter what's being talked about. That's exactly what an LLM is. No "consciousness," no "magic." Just a whole lot of reading, and then very, very good continuing-to-write.
Why is it called "Large"?
The "Large" refers to the model's parameters. Parameters are, put simply, the adjustment screws inside the AI's network. Each of these adjustment screws is a number. During training, each individual one of these numbers is gradually tuned so that the model gives good answers.
Early language models had a few million parameters. Today's LLMs have billions to trillions. If you read something like "Llama 70B" online, that means: this model has 70 billion parameters. GPT-4 is estimated at over a trillion, the exact number is a trade secret.
Parameters are not data. They are not the texts the AI was trained on. They are the traces that training left behind in the model, condensed knowledge, in the form of billions of small numbers.
More parameters = a better model?
Tends to be yes, but not always. A 7-billion-parameter model is smaller, faster, and cheaper to run than a 70-billion-parameter model, but it can reason less well, handles fewer languages, and makes more mistakes. On the other hand, there are small models that deliver really good results on specific tasks (e.g. writing code) because they were trained very precisely for them.
Size beats mediocrity. But smart training beats size.
How is a model like this trained?
An LLM doesn't learn "from a teacher." It learns from pure text. A whole lot of pure text. The training data consists of texts that are publicly available on the internet or have been licensed: Wikipedia, books, news, forum discussions, scientific articles, programming code.
The core exercise in training sounds almost trivial: the model is shown a piece of text, a word is removed, and the model has to guess which word is missing. At first it guesses randomly. But with every wrong prediction, the parameters get nudged a tiny bit, in the direction that makes the guess slightly better next time.
Sentence: "The sun rises in the ___." The model sees millions of sentences during training and learns: after "sun rises in the" comes almost always "east." Multiply that by every conceivable sentence beginning and every language, and you have a rough idea of how an LLM actually learns language over the months of its training.
One important point here: an LLM generates text piece by piece. Not letter by letter, but also not in whole words. It uses tokens, small building blocks that are sometimes a word, sometimes a syllable, sometimes a punctuation mark. If you want to know more about that, check out How does AI work?.
The best-known models
There are dozens of LLMs today. You'll hear the following six most often, good to know who's behind them.
GPT-4 / GPT-5
The models from OpenAI that power ChatGPT. For a long time the benchmark in language understanding. Closed source, you use them through the cloud.
Claude
The model family from Anthropic. Known for long context windows, good reasoning, and especially careful answers on sensitive topics.
Gemini
The model from Google DeepMind. Tightly integrated with Google Search and strong at multimodal tasks (text, images, video, audio at once).
Llama
The open model family from Meta (i.e. Facebook). Open source, anyone can download the weights and run them locally. Very popular in research.
Mistral
Models from the French startup Mistral AI. A mix of open and commercial. Known for high quality at a comparatively small model size.
DeepSeek, Qwen & Co.
Strong models from China (Alibaba, DeepSeek). Many are openly available and are shaking up the international landscape considerably, especially in math and code.
Open Source vs. Closed Source
An important distinction among LLMs: are the model weights, i.e. the billions of parameters, publicly accessible, or hidden away only in the provider's cloud?
Closed Source (e.g. GPT, Claude, Gemini)
- The model only runs in the provider's cloud.
- You talk to it through an interface (website or API).
- Generally the qualitatively strongest models.
- You pay per use. Your requests leave your system.
- No one outside the company knows exactly how the model is built.
Open Source (e.g. Llama, Mistral, Qwen)
- The weights are freely downloadable.
- You can run the model on your own computer or server.
- Your data stays with you, important for privacy.
- But you need powerful hardware, especially lots of graphics memory.
- You can customize the model and further train it on your own data.
"Open source" isn't quite the same for LLMs as it is for regular software. Often only the weights are published, not the training data or the exact training process. Some more accurately call this "open weights."
Context window, how much the AI holds in mind at once
An LLM has no long-term memory. It forgets everything after each conversation. What it can see and process right now is its context window, the slice of context that's fed into the model along with the current request.
This window is measured in tokens. Older models had 4,000 tokens, about 3,000 words. Modern models reach 200,000 up to 2 million tokens. Two million tokens correspond to roughly 1,500 book pages that the model can take in at once.
The context window is like the AI's desk. Everything you put on it, your question, uploaded documents, the conversation so far, the AI sees all at once. What doesn't fit on the desk doesn't exist for it. When the desk is full, something falls off the bottom end, usually the beginning of the conversation.
Even if a model has a huge context window: the fuller the window gets, the worse the quality often becomes. AIs "focus" on the beginning and the end of a long text. Content in the middle can easily get lost. This is called lost in the middle.
Multimodal LLMs, no longer just text
Classic LLMs only understand text. Modern models are multimodal: they also accept images, audio, and sometimes video as input, and can also generate images or speech.
You can show Claude or GPT-4 a screenshot, for example, and ask "What do you see here?" You can send Gemini a video and have it summarize it. This works because the models have learned to internally convert images and sounds into tokens too, into a language the model understands.
You photograph a handwritten shopping list. The LLM reads it, types it up neatly, sorts the entries by supermarket aisle, and suggests a recipe that could be made from the ingredients. Three steps, all within one multimodal model.
Mostly plausible, occasionally creative. Good for explanatory texts.
What LLMs are good at, and where they fail
Once you understand how an LLM learns, namely from patterns in texts, you also understand why it's brilliant at some things and surprisingly bad at others.
Strengths
- Rephrasing, summarizing, translating texts.
- Brainstorming ideas and drafting structures.
- A feel for language in emails, applications, advertising.
- Writing and explaining code in many programming languages.
- Translating complex topics into simple words.
Weaknesses
- Exact math, especially with large numbers.
- Counting (letters in a word, words in a text).
- Current facts that happened after training ended.
- Very rare specialist-knowledge questions, where hallucination is common.
- Genuine logic puzzles that didn't appear in training.
An LLM sounds just as confident with a wrong answer as with a right one. It has no built-in "I'm not sure." More on this in the chapter Hallucinations. Rule of thumb: always double-check important facts.
Why can't an LLM do math?
Because it isn't actually a calculator. It's a language predictor. When you type "17 × 23 =", the model isn't looking for the mathematically correct result. It's looking for the most likely next token. For simple problems that came up often in training, this works. For large numbers, it's a guess.
The modern solution to this: the LLM gets a calculator handed to it as a tool, through what's called tool use or functions. More on this in the chapter AI agents.
An LLM doesn't know what's true. It knows what sounds like it's true. That's an important difference.
Who actually does what?
Three terms are often confused, but belong together:
- The model, e.g. GPT-4, Claude Sonnet, Llama 70B. The raw weights.
- The product, e.g. ChatGPT, claude.ai, Gemini. The chat interface you see as a user.
- The API, the access point developers use to build the model into their own software. More in API & SDK.
ChatGPT is not GPT-4. ChatGPT uses GPT-4 (or another model). Likewise, claude.ai is the interface. Claude Sonnet, Opus, or Haiku are the actual models behind it.
What you now know
- An LLM is a massive neural network with billions of parameters that has learned from texts to predict language token by token, hence the "Large."
- Models differ in size (parameters), openness (open vs. closed source), field of view (context window), and whether they process only text or also images and sounds.
- LLMs excel at language, summarizing, and ideas; for exact math, counting, and up-to-date facts, they need to be given tools or sources.