Definition
A proprietary model is one whose owner materially controls access to its weights, development artifacts, operation, modification, or redistribution. Users commonly receive access through a hosted API or application. Some vendors distribute weights under licenses that retain enough restrictions or closed components for the larger system to remain proprietary.
"Proprietary" describes control, not one packaging model. A provider may publish architecture papers and evaluations while withholding base weights. Another may permit vendor-managed fine-tuning or release weights that cannot be redistributed. The label alone does not reveal which artifacts are closed or which rights are restricted.
Operational significance
Using a proprietary model can transfer serving, optimization, patching, and some abuse controls to the provider. The same arrangement creates dependencies that a factory must govern:
- model versions or behavior may change behind a stable API name;
- prices, rate limits, regions, retention rules, and product availability can change;
- internal weights and training data cannot usually be independently inspected;
- reproducibility may depend on provider-controlled snapshots and inference infrastructure;
- sensitive prompts, context, or outputs cross an organizational boundary unless a dedicated deployment arrangement says otherwise;
- exit costs rise when prompts, tools, evaluations, or application logic depend on provider-specific behavior.
The decision depends on the workload. A managed service may provide stronger operations or lower cost at one scale, while a self-hosted model may be necessary for another data boundary. The label does not settle security, quality, or total cost.
Distinguish it from nearby terms
Closed-weight specifically means the trained parameters are unavailable. Open-weight describes access to those parameters, not a complete grant of open-source freedoms. Open-source AI requires broader artifacts and rights. Frontier describes capability position, so a model can be proprietary without being frontier or frontier without being proprietary.
Check your understanding
An API provider can change a model behind a stable alias and offers no immutable snapshot. Which evaluation, version, retention, and exit controls does a production system need before it can depend on that endpoint?