Skip to content
atlas

Don't confuse these

Large language model (LLM) vs Small language model (SLM)

Why they differ

Both work the same way, but a small language model has far fewer parameters, so it is cheaper and faster but knows less.

Large language model (LLM)

Language models

A very large model trained on huge amounts of text to predict the next word, which lets it write, sum up and answer in fluent language.

Formal

A deep learning model, usually a transformer with billions of weights, trained on vast text collections to predict the next token; the same skill, repeated, lets it produce long answers to a prompt.

In plain English

Like the next-word suggestions on a phone keyboard, but trained on a large part of the internet - it knows how answers usually sound, not whether they are true.

In practice

Staff in a municipality get a chat assistant built on an LLM to draft replies to citizens and sum up meeting notes, and the IT department must decide which data they may paste into it.

Why it matters

LLMs sound confident even when wrong, can be steered by hidden text, and send your input to whoever runs them - all real risks to weigh before use.

Small language model (SLM)

Model architectures

A language model small enough to run cheaply on a laptop or phone, trading some broad skill for speed, privacy and low cost.

Formal

A transformer built like a large language model but with far fewer model parameters - from a few hundred million up to around ten billion - often trained on carefully chosen data or through knowledge distillation from a larger model.

In plain English

Like a pocket dictionary next to a full encyclopedia - it will not know everything, but it fits in your bag and answers the common questions right away.

In practice

A home-care nurse in a municipality dictates visit notes into a tablet that runs a small language model; it drafts the record entry even where there is no mobile signal, and nothing leaves the device.

Why it matters

For narrow, repeated tasks it can match a bigger model at a small share of the cost and delay, and keeping data on the device makes data protection far simpler.

Shared connections

Atlas is in beta.