The tech
A language model is software trained on huge amounts of text to predict the next word in a sequence. That simple skill is what lets it draft, summarise, translate and answer questions.
Language models build on natural language processing (NLP), the branch of AI concerned with understanding and generating human language. NLP also covers narrower tasks, such as identifying the topic of a message or pulling names out of a document.
Models produce text one token at a time. A token is a word or part of a word. What a model learns during training is stored as parameters: numerical values that shape each prediction. The number of parameters is the main difference between the two broad types:
- Large language models (LLMs) have more parameters and handle a broad range of tasks, but need more computing power to run.
- Small language models (SLMs) have fewer parameters, so they suit settings where memory, computing power or budget is limited. They can be built small from the start or "distilled" from a larger model.
There's no agreed size boundary between the two. Microsoft, for example, calls its 14-billion-parameter Phi-4 a small language model.
Size is also separate from openness. A small or a large model can have open or closed weights (see Open weight models).
Why it matters
- Not every task needs a language model. Sorting messages into fixed categories may only need a conventional NLP system. Flexible drafting, or answering varied questions, is where a language model earns its place.
- Bigger isn't automatically better. Larger models offer broader capabilities, but training data, model design and the specific task matter as much as size. A well-trained small model can beat a larger one on a narrow job.
- Smaller models are cheaper and easier to run. SLMs need less memory and fewer graphics processing units (GPUs), which lowers hardware, energy and running costs. That matters for organisations without large server budgets, and in places where the power supply is constrained.
- Some can work offline. The smallest models can run on a laptop or a recent smartphone without an internet connection, which opens up use in the field where connectivity is patchy.
- Fluent isn't the same as correct. Models of every size can produce plausible but wrong answers, and can misread slang, sarcasm or local context. Performance is often weaker in languages that had little training data, so it needs testing in each language you plan to use.
In the real world
Potential applications include:
- Processing community feedback. NLP tools could categorise survey responses, spot recurring concerns or extract place names, with checks for differences across languages and dialects.
- Working with limited connectivity. A suitable SLM could run locally to help staff search guidance or draft routine text, where devices have enough capacity.
- Working across complex documents. An LLM could help synthesise programme reports or draft briefings, with sources checked and a person reviewing the output.
The choice should follow testing on representative local data. Compare task accuracy, performance in the languages that matter, running cost and staff effort. A smaller model that does a defined task reliably may be more useful than a larger one that's hard to run or maintain.