Definition
There is no fixed size cut-off, but small language models typically range from under one billion to a few tens of billions of parameters. Examples include Microsoft's Phi family, Google's Gemma and smaller Llama, Mistral and Qwen models.
SLMs trade breadth for efficiency. On a well-defined task, such as classification, extraction or routing inside an agent, a fine-tuned small model can match much larger ones at a fraction of the cost, run on-premises for data control and respond faster. NVIDIA researchers have argued that small models are the better default for many repetitive agent tasks.