What is a small language model (SLM)?
An SLM is a much smaller language model that can run even on a phone or laptop. It is fast and cheap for focused tasks.
What is an SLM?
A small language model is the lightweight sibling of large language models (LLMs). There is no strict cutoff, but models from a few hundred million to a few billion parameters usually go by this name, while the largest models reach hundreds of billions. Microsoft's Phi, Google's Gemma and the small versions of Meta's Llama family are well known examples.
Why small?
A small model needs less memory and less computation. That means it can run on a phone, a laptop or a company's own server instead of the cloud. Data stays in house, which makes privacy requirements easier to meet. Inference costs drop, answers come back faster and latency falls. Some small models can even work without an internet connection.
What are they good and bad at?
SLMs are strong at narrow, well defined tasks: classifying email, summarizing text, extracting data from forms and simple conversations. Adapted to a specific job with fine-tuning, they can come close to large models on that job.
Broad general knowledge, multi step reasoning and long, complex documents remain the territory of large models. Small models make more mistakes on such tasks and are more prone to hallucination.
How businesses use them
A common approach is to route work by difficulty: a fast, cheap SLM handles most incoming requests and the hard ones go to a large model. That routing happens in the orchestration layer, keeping quality up while total cost goes down.