SLM (Small Language Model)

Technical terms
About 1 min read

An efficient model with billions of parameters that delivers high performance even on low-spec hardware, optimized for on-device AI and task-specific specialization.

Also known as
Small Language ModelLightweight Language ModelSmall-scale Language Model

Detailed explanation

An SLM is an artificial intelligence model designed to improve computational efficiency by optimizing the number of parameters to a range of hundreds of millions to tens of billions (typically 1B to 10B, and broadly under 30B). Unlike Large Language Models (LLMs), SLMs can run independently on individual devices (on-device) such as smartphones and laptops without a cloud connection, offering ultra-low latency and low operational costs. Modern SLMs are built using high-quality, curated training data or knowledge distillation techniques from larger models, enabling them to deliver LLM-grade performance on specific business logic or specialized domain tasks. Since data is not transmitted externally, they offer excellent security and play a critical role in privacy-sensitive environments like finance and healthcare, as well as in building offline AI agent systems.

Why it matters in tool selection

SLMs make it possible to commercialize AI capabilities without heavy infrastructure. They are a much more logical choice than LLMs, especially for businesses looking to drastically reduce API call costs, or those that need to run AI on-premise or in local environments for data privacy.

What to look for

  • Support for Quantization: Checking whether the model can be loaded into mobile device memory when compressed to 4-bit, etc.
  • Inference speed (Tokens per second): Measuring whether the target response speed is achieved in actual hardware environments (CPU/NPU)
  • Fine-tuning efficiency: Verifying whether task performance can be sufficiently improved even with a small amount of domain-specific data

Representative SLMs

Representative examples include Microsoft's Phi-3-mini (3.8B), Google's Gemma 2B/7B, Meta's Llama 3.2 (1B/3B), and Mistral's Ministral 3B/8B. These models run smoothly on iPhones or standard laptop GPUs.

Related terms

LLMOn-device AIFine-tuning