TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Generative Mechanics · AGE 16—18 · FACTUAL · EVOLVING
Small models, large models
Size is one side of a trade, and size alone does not rank models.
The idea
Models come in sizes. Size means, roughly, how many internal numbers the model carries. They are called parameters. Larger models often do more kinds of task. But size alone does not rank models: how a model was trained, whether it routes each token through only part of itself, whether it was distilled or shrunk for speed, and how much compute it spends at answer time can all overturn a size comparison. Larger models cost more and live on someone else's servers. Smaller models run on a laptop and keep your material at home. Size is one side of a trade.
Why it matters
The useful question is never "which is the best model?" It is "which size fits this job, this budget, and this material?"
See it in the studio
Tagging ten thousand site photographs by building type is a narrow, repeated job. A small model on a studio machine can do it overnight, with no fee and no upload. Reasoning through a contradictory client brief is a wide job. That is where the large model earns its cost.
Watch for this
Reading a model's size as a promise of accuracy. Large models make confident mistakes too, and they write them more fluently.
Try it
Find the parameter count of two models you have heard of — many are public. Give both the same simple task: name the style of a building in a photograph. Note speed, cost, and whether the bigger one was actually better at this. If it was not, you have found the trade for yourself.
Prove it
Name one task where a small model is the right choice and one where a large one is, and give the trade that decides each.
How it works
Parameters are the weights adjusted during training. More of them means more room to store patterns, more memory to hold them, and more computation per answer. That is where the cost, the delay and the energy bill come from. "Size" can mean several things, though: total parameters, the parameters active for one token (mixture-of-experts models use only a part of themselves per token), memory footprint, or compute per answer — and two models of the same count can differ a great deal in speed and in what they do well. Techniques that shrink models keep changing what a small one can do, and new large models keep raising the top. So the trade is real, and the numbers move. That is why this card is EVOLVING.
What this idea builds on
- A starting idea.
What this idea opens up
Sources
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.