LLM (Large Language Model)
A large language model is a system trained on very large amounts of text to predict the most likely continuation of a sequence. Everything it appears to do — summarising, drafting, answering, translating — is that one operation applied to different inputs. It holds no facts in the way a database does, which is why the useful enterprise question is never "what does it know" but "what was it shown".
Why does the prediction framing matter?
Because it explains the failure mode. A model produces the continuation that fits best, not the one that is true. A fabricated case citation and a real one are formally indistinguishable to it — both look like the kind of thing that follows that kind of sentence.
It also explains why more parameters do not fix the problem. A larger model predicts better, so it is wrong less often; the mechanism that makes it wrong is unchanged. In work where a statement has consequences, less often is not a category.
What separates models in practice?
For enterprise work, less than the benchmarks suggest. Where the answer has to come from documents you supply, the model is reading rather than remembering, and the gap between a frontier model and a good open-weight one narrows sharply.
What does separate them: how well they follow instructions under pressure, how reliably they produce structured output, how they behave when the answer is not in the material, and how much they cost per million tokens at your volume. The last two decide more procurement arguments than capability does.
Which model should a company pick?
The one you can change later. The binding decision is not which model but whether your documents, retrieval and system instructions are independent of it. Teams that separate these switch models in days; teams that do not rebuild.
That is also the honest answer to "will this be obsolete in a year". The model will be. The corpus, the retrieval and the workflow around them will not.
What it is not to be confused with
Chatbot
A chatbot is one interface to a model. The model can also sit behind a search box, a document pipeline or a batch job with no conversation anywhere. Choosing the chat interface by default is the most common avoidable mistake in enterprise deployments.
Search engine
A search engine returns documents; a model returns prose. Combining them — retrieval for the documents, the model for the prose, citations for the link back — is the design most enterprise use cases actually want.
Frequently asked
What does LLM stand for?+
Large Language Model. Note that the abbreviation collides with Master of Laws in legal contexts, which is why search results for it are often about postgraduate degrees rather than machine learning.
Do we need the largest model available?+
Rarely. Where the system answers from documents you supply, a mid-sized open-weight model is frequently indistinguishable in output and several times cheaper per token. The place to spend on capability is reasoning over many documents at once, not everyday drafting.
Can a model be trained on our company data?+
It can, and for most organisations it should not be — at least not first. Training raises purpose-limitation questions under the GDPR, has to be repeated as data changes, and produces a model that cannot cite its sources. Retrieval gives you current information with citations and no change to the model.
What is a token?+
The unit a model reads and writes — roughly three-quarters of a word in English, less in German where compounds split. Pricing and context limits are counted in tokens, which is why a German document costs more to process than an English one of the same length.
The procurement question worth asking
Ask a prospective supplier what changes if you switch the underlying model in six months. The answer separates people who built an application from people who wrapped an API.
Request a test