Small Language Models are language models with a reduced number of parameters, typically between a few hundred million and a few billion, compared to the hundreds of billions in large models (LLM). This smaller scale makes them lightweight, fast, and efficient, capable of running locally on laptops, mobile phones, or edge devices without depending on cloud servers.
Their relevance lies in several practical factors:
- Privacy: data can be processed on the device itself, without being sent to third parties.
- Cost and latency: they require fewer computing resources and respond more quickly.
- Sustainability: they consume less energy during training and inference.
In exchange, they often offer more limited performance in complex tasks or broad knowledge. Therefore, many SLMs are specialized or fine-tuned for specific domains, where they match or outperform larger models. Well-known examples include the Phi families from Microsoft, Gemma from Google, or Llama in its smallest versions.