Open-Weight vs Closed AI Models
The key question is not whether a model feels open. It is which artifacts you can access, what the license permits, and where inference can run.
Start with precise terms
A closed model is typically accessed through a provider-controlled product or API without distributing the model weights to the user. An open-weight model makes trained weights available for download under some set of license terms.
These categories describe access to an important artifact, but they do not by themselves describe the complete development process, training data, source code or legal permissions.
Weight access changes what you can operate
With downloadable weights, you may be able to run inference on your own hardware or infrastructure, choose a compatible runtime, quantize the model, inspect its architecture and preserve a particular model version. Hardware requirements still determine whether a given model is practical locally.
With a closed hosted model, the provider operates the inference infrastructure. This can remove local hardware requirements and simplify access to large systems, but model availability, pricing, rate limits and behavior can change with the service.
Deployment affects privacy and control
Local or self-hosted inference can keep prompts and model execution inside infrastructure you control, subject to the rest of your software stack. Hosted inference sends requests to an external service and should therefore be evaluated using that provider's current data-handling and retention terms.
Neither deployment style is automatically secure. Local systems still need access control, updates and operational security; hosted systems still need careful credential, data and permission management.
Open weights does not automatically mean open source
The phrase open source AI has legal and technical implications beyond the simple ability to download weights. An open-weight release can impose use restrictions or omit training code and data. Conversely, a repository containing source code does not necessarily grant unrestricted rights to model weights.
Do not infer commercial, redistribution or modification rights from the words “open model.” Read the license attached to the specific release you intend to use.
Choose based on the operational requirement
Open-weight models are especially relevant when offline operation, self-hosting, reproducibility or control over a specific model artifact matters. Closed services can be practical when you prefer managed infrastructure or need capabilities available only through that provider.
Treat this as an architecture decision rather than an ideological label: identify the required capability, data boundary, deployment environment, budget and license constraints, then evaluate candidate models under those conditions.