TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Foundations · AGE 16—19 · FACTUAL · EVOLVING
Open and closed models
A model's release decides whether your uploads leave your building and what each use costs.
The idea
Some models are released as files you can download: the weights. Anyone with the hardware can run and fine-tune them. Others are reachable only through a company's API, and the weights stay private. That decides whether your uploads leave your building, what each use costs, who can change the model. Four overlapping categories, not two: proprietary and hosted; open-weight, where the files are public but training data and code may not be; open-source AI under the Open Source Initiative's definition, which needs code and training-data information too; and locally runnable. Most "open" models today are open-weight, not open source.
Why it matters
Control, privacy and cost follow from how the model was released, not from how good it is. Ask "open weights or API?" before asking "is it good?"
See it in the studio
A practice with confidential drawings can run an open-weights model on a machine in the office and never send a plan outside. With an API model the plan travels to the provider's servers under the provider's terms. That may be fine, but it is a decision someone must take.
Watch for this
"Open" used as a brand. A downloadable model can come with a licence that forbids commercial use or particular uses. Read the model card and the licence. The word in the headline is not the same as the terms.
Try it
Pick two AI tools you know. For each, find out: can the weights be downloaded? Under what licence? If not, whose API does it call? Write it as a two-row table. One row will probably be harder to fill, and that is information too.
Prove it
Explain the difference between an open-weights model and an API-only model, and name one thing each choice makes easier and one it makes harder.
How it works
A model's behaviour lives in its weights, the trained numbers. Releasing them lets others run, fine-tune, test and deploy the model offline — though reading billions of numbers is not the same as reading a program's logic, so "inspect" promises more than it delivers. Keeping them private lets the maker control use, charge per call, and update the model without telling you. Hosting sites such as Hugging Face carry the open ones with model cards stating intended uses, training data and limits. EVOLVING because the licences, the definition of "open", and which makers release what are all changing quickly. The Generative Mechanics strand carries the fuller structural decision for a practice.
What this idea builds on
What this idea opens up
Sources
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.