A compact AI language model, often a fraction of the size of a frontier LLM, built to run cheaper, faster, and often locally, by trading general purpose breadth for strong performance on a narrower set of tasks.
Also known as SLM
Small language models are built the same way large language models are, but with far fewer parameters, sometimes a few billion instead of hundreds of billions, and often trained more narrowly on the tasks they actually need to perform. The trade-off is real: an SLM will not match a frontier model’s breadth of knowledge or ability to handle an open ended, unfamiliar request. But for a well defined, repeatable task, a well tuned small model can perform close to a much larger one at a fraction of the cost and latency.
2026 is being described as the year efficiency caught up with capability. Enterprise spending on local, on device model execution rose sharply as teams realized most day to day AI tasks do not need a frontier model’s full power, they need a fast, cheap, reliable one running on infrastructure the company already controls.
Most agency AI use is not asking a model to write a novel, it is running the same kind of task hundreds or thousands of times: tagging assets, drafting ad variants, summarizing briefs, classifying feedback.
Cost adds up fast at agency scale. Running a frontier model for every routine, repeatable task across dozens of clients gets expensive quickly, a small model tuned for that specific task can deliver comparable results at a fraction of the per call cost.
Speed matters for production workflows. A small model responding in milliseconds instead of seconds keeps tools like real time ad copy variation or live chat responsive in ways a slower, larger model cannot.
Data can stay local. Some SLMs run entirely on a client’s own infrastructure or device, which matters for clients in regulated industries who do not want data leaving their environment to reach a cloud hosted frontier model.
Your agency runs a content operation that tags and categorizes thousands of client social assets a month, sorting by product, campaign, sentiment, and compliance risk. Running that volume through a frontier model would be accurate but slow and expensive at scale. Instead, your dev team fine tunes a small language model specifically on the tagging task, using a few thousand examples of correctly labeled assets. The resulting model runs the classification in a fraction of the time at a fraction of the cost, freeing frontier model budget for the harder, more open ended creative strategy work where that extra capability actually earns its keep.
The automations and agents module of the workshop teaches you how to build AI workflows that compress the busywork without taking the craft out of the studio.