A language model running on far fewer parameters than an LLM, focused on a narrow set of tasks and cheap enough to run on a single device.
What is an SLM?
SLM stands for Small Language Model. It shares the basic architecture of an LLM but runs on far fewer parameters. The consequences are practical: it answers faster, costs less to run, and often fits on a single device rather than a server fleet.
What it gives up is generality. You do not ask an SLM for depth on any subject; it is good at bounded, repetitive work — classification, tagging, short summaries, format conversion, intent routing. "Small" here means narrow, not inadequate.
Where SLMs are useful
The right question is not which is better but which the job requires. Sending every task to the largest available model is like renting a truck to deliver an envelope: it works, but it is needlessly expensive and slow.
- High-volume repetitive work — classifying reviews, tagging products, parsing forms — is solved far more cheaply by small models.
- Where latency is critical, such as on-device features or instant suggestions, a small model becomes a requirement rather than a preference.
- When data must not leave the device, running locally is an advantage in itself.
In practice the two model sizes are not rivals but a division of labour: the small model handles routine steps, the large one handles the step that needs reasoning. The most visible form of that arrangement is AI agents.
Frequently asked questions
What is the difference between an SLM and an LLM?
Scale and purpose. An LLM covers a very wide range of knowledge and capability, at higher cost and latency. An SLM runs on far fewer parameters, trading generality for speed, cost and the ability to run on-device. For narrow, repetitive work an SLM fits; for open-ended reasoning an LLM does.
Does using a small model mean sacrificing quality?
Not if the task is chosen well. On bounded work like classification, tagging or format conversion, a small model can come close to a large one at a fraction of the cost. The trade-off starts with open-ended, deep reasoning — where a large model is still required.
Related concepts
AI Agents
A system where a language model plans and carries out a multi-step goal by using tools; it does not just produce text — it searches, reads pages and takes actions.
LLM (Large Language Model)
An AI model trained on very large volumes of text that generates language by predicting what comes next; the technology behind products like ChatGPT, Claude and Gemini.