Large language model (LLM)
A large language model is a system trained on very large amounts of text to predict what text comes next. That single capability, applied at sufficient scale, produces the ability to answer questions, summarise, classify, and write in a given style.
A large language model is a component rather than a product. Every AI chatbot, agent, and assistant is built on one, and the model supplies language and reasoning while everything else (content, tools, permissions, and checks) is built around it. Comparing vendors by which model they use usually misses where the real difference sits.
How does a large language model work?
Training exposes the model to enormous quantities of text, from which it learns statistical relationships between pieces of language. Nothing is stored as facts in a retrievable form. What remains is a very large set of weights encoding how language tends to fit together, and that design explains most of the behaviour teams find surprising:
It does not know what it knows. There is no index to check, so a gap and a fact feel identical
It has a training cut-off. Anything after that date is absent unless supplied at the time of the question
It has no access to your systems unless something connects it to them
It is probabilistic. The same question can produce different wording, and occasionally different substance
Those four are why production systems wrap a model in retrieval, tools, and checks rather than using it directly.
What can a large language model not do on its own?
On its own a model can only produce text. It cannot look anything up, take an action, or verify what it produced, so a deployment consisting of a model and a prompt will answer confidently and generally about matters it has no information on.
In financial services that gap is most of the job. A customer asking about their account needs the account record, and a customer asking for a refund needs something able to issue one. The model handles the conversation, and the system around it decides whether anything actually gets resolved. Our guide to evaluating AI agents covers how to test that difference.
Frequently asked questions
What is the difference between an LLM and generative AI?
Generative AI is the broader category, covering systems that produce images, audio, video, and code as well as text. A large language model is the text and language member of that family. Most business applications people call generative AI today are built on language models.
Do we need to know which LLM our vendor uses?
Less than most procurement processes assume. Model choice changes often, and a vendor tied to one today may not be in six months. The durable questions are which providers process your data, where, whether you are told when that changes, and whether your evaluation results still hold afterwards.
Are larger language models always better?
No, and cost and latency usually settle it. A larger model reasons better on hard cases while running slower and costing more on easy ones, which is why production systems often route simple tasks to smaller models. Accuracy on your own cases matters more than any general benchmark.