A small language model (SLM) aims to deliver useful language capabilities with less compute, memory or serving cost than a larger model. Smaller models can be attractive for classification, extraction, simple assistants and on-device use, especially when tasks are narrow and evaluation is clear.
Typical fits include routing a message to the right team, extracting fields from invoices, or running an assistant on a phone without a network connection. For open-ended reasoning over long documents, a larger model may still do better. A sensible approach is to try the small model on a sample of real tasks and move up only where it fails, which is also the logic behind model routing.
Small does not always mean faster or cheaper in every setup; runtime efficiency and hardware matter. Compare quality on real examples, including difficult cases, before replacing a larger model. Related: Model Quantization, Knowledge Distillation.