What does Large Language Model mean?

A large language model (LLM) is a language model with several billion parameters that continues text by predicting and appending the next token, one step at a time. It is trained on very large volumes of text without the later tasks being fixed in advance. It does not store documents verbatim, but statistical relationships between sequences of characters — hence the name language model.

The transformer architecture relates every token to all others in the context and weighs these relationships through self-attention. Pre-training is followed by two further steps: adaptation to example tasks, and alignment with human ratings. The context window limits how much text a model can take into account at once, today typically between 8,000 and several hundred thousand tokens. Models with open weights can be downloaded and run independently; closed models are only accessible through the provider’s interface.

An LLM pays off wherever language is the material being worked with and the tasks change from case to case. Typical tasks are summarising, translating, answering questions about an organisation’s own documents, and converting free text into structured fields. For calculations and exact lookups, specialised applications and databases remain more reliable.

Its advantage over specialised models is that a single model covers many tasks and requires no training of its own. It is operated in ordinary language, so staff without technical background can work with it. Existing specialist systems remain untouched, because the model runs alongside them rather than inside them.

An LLM only knows what existed at the time it was trained, and fills gaps with plausible-sounding inventions. For current and organisation-internal content, it therefore needs controlled runtime access to verified sources, usually through a connected document search.