Open-Weight AI Models

An open-weight AI model makes its trained numerical parameters, called weights, available for users to download and run. A proprietary hosted model keeps those weights under the provider's control and offers model capabilities through an application or API. Open-weight describes access to model parameters, while hosted describes the method of providing access; they are not perfect opposites.

What access to weights enables

Weights encode the parameter values learned during training. Their availability can permit users, subject to licence and technical constraints, to deploy locally, modify or fine-tune the model, conduct independent tests, and integrate it without sending every query to the original provider.

• Local operation can provide greater control over data handling, model configuration and continuity of access.
• Users must supply suitable computing infrastructure, security controls, updates and technical expertise.
• Independent examination is easier than with an API-only system, but weights alone may not reveal the training data, training process or complete source code.

How a proprietary hosted model differs

In a hosted arrangement, the provider operates the model on its own infrastructure and users obtain outputs through an API or interface. The provider controls the underlying weights, deployment environment, safeguards, pricing and model updates.

• Users generally cannot inspect, copy or independently modify the underlying parameters.
• Hosting reduces the user's infrastructure burden, but creates dependence on the provider's access rules, service continuity and data-governance arrangements.
• The provider can change or withdraw the model without distributing its internal parameters.

Open-weight is not necessarily open-source

Availability of weights does not automatically make a model open-source. Its licence may restrict commercial use, redistribution or particular applications, while training code, datasets and documentation may remain undisclosed.

• A model's openness must therefore be assessed across weights, code, data, documentation and licensing rather than by a single label.
• An open-weight model may also be offered as a hosted service, so parameter access and delivery mode should be examined separately.

How UPSC asks this