Definition
A large language model (LLM) is a neural language model trained at substantial scale. During training, it learns patterns in token sequences. At inference time, it applies those learned patterns to the tokens in its input context.
Scale can help a model learn broad language capabilities, but it does not guarantee them. Training choices and later adaptation matter too.
“Large” refers to scale relative to other language models, not to a fixed parameter count. Transformer architectures are common among LLMs, but the architecture is not part of the definition. Many also process source code because code can be represented as token sequences alongside natural language.
Simple example
The same instruction-tuned LLM might summarize an incident report and then extract its service name and severity into JSON. The application changes the task through instructions and context instead of training a separate model for each job.
That flexibility does not guarantee equal performance across the tasks. A model that writes a convincing summary may still omit a relevant detail or produce invalid structured data.
Why it matters
A capable general-purpose LLM can support several language features without requiring a separate task-specific model for each one. This can reduce the amount of additional training needed for an initial implementation.
But fluent output is still a model prediction. A model can check its own output or follow a policy, but those responses are probabilistic too. They do not guarantee correctness or compliance. Teams need to evaluate the model on their own tasks. Hard rules need controls outside unconstrained model output. Model and request size can also change runtime requirements and cost.
One important nuance
There is no universal threshold at which a language model becomes “large.” The label changes with the state of the field and says little about whether a model is suitable for a particular task.
LLM and foundation model describe different properties. LLM describes a language model primarily in terms of scale. Foundation model describes a model’s role as a reusable base for downstream applications. Many LLMs are foundation models, but a foundation model can work primarily with another modality.