Small language model (SLM)
A language model trained for one domain and task: it reads like your expert, calls the right tool, keeps the case's memory, and writes in your format.
Category
How a model is chosen and adapted
Related terms
An SLM is a language model trained for one domain and one task: it reads and reasons like your expert, calls the right tool with the right parameters, keeps the case's memory, and writes in your format. On that task it often matches a large general model. “Small” is relative and moves every year; the hardware is sized for the task, never assumed.
Why size is a design decision
A large model is not automatically the right answer to an institutional task. Task scope is usually narrow: read one document class, produce one structured output, call one set of approved tools. A model that fits that scope can be run where the data already lives, evaluated more cheaply, and versioned without renegotiating a provider contract.
Size trades against generality. A large model is more forgiving of vague prompts and unfamiliar phrasing. A small model needs the task, the vocabulary, and the output contract to be defined. That definition work is where most of the reliability comes from anyway.
When an SLM fits
The signals are: the task is bounded, the document class is stable, the output is structured, latency matters, the data cannot leave the environment, or per-call cost is a real constraint at production volume. Several of those together make the case strong.
The signals against: the task is open-ended, the input is unpredictable, or the required reasoning depth has not been demonstrated on a representative evaluation set at the smaller size.
What does not change
A smaller model does not relax the controls around it. Policy, identity, permissions, delegated authority, and stop conditions stay outside the weights regardless of model size, and the model is evaluated against the same acceptance gates. An SLM that fails a representative evaluation is not shipped because it was cheaper to run.
Often confused with
How it differs from similar terms.
| small language model (SLM) | Compared with | What's different |
|---|---|---|
| SLM | Large language model (LLM) | Chosen for a bounded task, private deployment, latency, or cost. Needs the task and output contract defined. |
| Model size | System capability | Capability comes from the system — retrieval, tools, policy, evaluation — not from parameter count alone. |
Want these terms worked through on one of your own decisions? Start with a Decision Blueprint