Large Language Model (LLM)

A large language model, or LLM, is a model trained on extensive data to process and generate language.

During text generation, it calculates probabilities for the next token step by step. Training, model architecture and further adaptation enable tasks such as summarising, translating, answering questions and working with code.

How an LLM gets its capabilities

Training typically involves several stages. During pre-training, the model learns statistical patterns from large datasets.

Further training and alignment stages adapt its behaviour to tasks, instructions and desired safety requirements. The exact approach differs between models and providers.

Access, licensing and operating model

LLMs differ in whether model weights are available, which licence terms apply and where the model can be operated.

Some offerings are available only as managed services. Others provide weights under different, sometimes restrictive licences and can be run on owned or rented infrastructure.

The full operating model matters for privacy and information security. This includes processing location, logging, retention, subprocessors, access controls and contractual commitments.

Self-hosted models provide additional technical control, but also transfer full operational and security responsibility to the organisation.

Where the technical limits are

An LLM generates output from learned statistical patterns and does not automatically verify statements as true. It can therefore produce convincing but incorrect content.

RAG, tools and supplied context can add current information, but do not eliminate the risk of hallucination.

The compute, latency and cost of inference depend on factors including the model, context length and usage.