Open vs. closed models

One building lets you in the door. The other only opens a window — and which one you're using decides what you're actually allowed to do with it.

A heads-up for this module: pages 22–27 use some real industry terms (like "API," just below) you'll meet outside this course too — you don't need to memorize every one on first read.

Closed

You send a request through that window — usually called an API — and get an answer back. The weights stay inside; you never touch them.

Open

You take a copy of the actual numbers home — to run as-is, change, or put on your own machine.

The word does three jobs at once

"Open" bundles three separate questions

Pick a release pattern below — watch which of the three questions it actually answers "yes" to (the third, licence freedom, just means what you're legally allowed to do with it).

Weights

—

Training data

—

Licence freedom

—

Pick a pattern above.

What actually drives most people's choice

Cost and capability, not just auditability

CostClosed bills you per request. Open hands you the weights for free — running them well is the part you now pay for, in hardware or rented compute.

CapabilityClosed models usually lead at the frontier; open ones trade some ceiling for control, and the gap between them keeps shrinking, not staying put.

Durability note: the shapes of these tradeoffs are stable — exactly how big the gap is right now isn't, so no numbers are worth writing down here.

So what

When someone calls a model "open," ask which of the three they mean — the answers to all three are almost never the same.