Large Language Model (LLM)

How it works

An LLM is trained on very large amounts of text and learns statistical patterns of language along the way. When you send a request, it calculates piece by piece which chunk of text (token) is most likely to come next — that is how whole answers emerge. It only “knows” things to the extent that they are reflected in these patterns. Well-known model families include GPT, Claude, Gemini, Llama and Mistral.

A practical example

You give an LLM a customer’s email and ask it to draft a polite reply. The model writes a text that fits the request and the tone — without you having to write rules for every individual case.

What you should know

• An LLM is not a reference book: it can be convincingly wrong (see hallucination).
• Its knowledge ends at its training cutoff. Current or internal information has to be provided to it, for example via RAG.
• Models differ greatly in quality, speed, cost and data protection. Which one fits depends on the task.
• Many LLMs are also available with open weights to run yourself. A selection with benchmark results and hardware needs is in our open AI models comparison.