← AI Terminology
SLM - Small Language Model
A small language model (SLM) is a compact LLM (often 1–10B parameters, sometimes smaller) designed for efficiency, edge deployment, or specialised tasks rather than frontier generality.
Distillation and careful data make SLMs surprisingly strong.
Distillation and careful data make SLMs surprisingly strong.
Why It Matters in AI
Not every task needs a frontier API. SLMs cut cost and latency, run on-device, and specialise via fine-tunes. The 2024–2026 wave of strong small open models reshaped product architecture.
Key Points
| Aspect | Description |
|---|---|
| Size | Roughly sub-frontier; mobile to single-GPU server |
| Pattern | SLM locally + frontier fallback |
| Related | On-device AI, distillation, GGUF |
| Examples | Phi, Gemma-2B/7B, Llama-3.2-1B/3B class models |
| Strengths | Cost, speed, privacy, specialty finetunes |
| Weaknesses | Hard reasoning unless carefully trained |
Simple Analogy
A nimble city car instead of a semi-truck — not for every cargo load, but perfect for daily trips and narrow streets.
Common Usage Examples
ollama run phi3- Mobile NPUs running SLMs
- Distill teacher into 3B student
- Classify intents with SLM; escalate hard cases
Summary
In short: SLMs are compact language models optimised for cost, speed, and edge use — often enough when frontier scale is overkill.