Open-weight models

What is an open-weight model?

An open-weight model makes its trained parameter values available under a licence. Those values are the numerical state used during inference. Access to weights gives developers a way to run a model on suitable infrastructure, instead of relying solely on a particular hosted endpoint.

Open weights and open source are related terms, but they do not mean exactly the same thing. A release may provide weights without providing all training data, training code or the details required to reproduce the model. The permissions attached to a release also matter. Availability of a download alone is not a statement about how it may be used.

What should a builder evaluate?

Start with the task: writing code, answering questions, processing images, generating speech or retrieving documents. The model card should describe relevant capabilities and limitations. Confirm the input and output formats rather than assuming that models with similar names accept the same requests.

Evaluate the exact version you intend to use. Instruction-tuned variants can behave differently from base models, and separate releases may change the prompt format or supported features. Keep a record of the model version, serving configuration and evaluation prompts so another team member can understand how a decision was made.

The model licence deserves its own review. Check permitted uses, redistribution conditions and any restrictions affecting the application. A capability demonstration and a licence review answer different questions; neither replaces the other.

Weights or a managed API?

Running weights yourself adds serving responsibilities. The team must select compatible software, allocate hardware, handle requests and monitor the deployment. A managed API exposes the model through a service interface, so application developers can focus more of their work on integration and evaluation.

The choice depends on the workload and operating model. Some teams need dedicated capacity and a particular deployment configuration; others want a straightforward API for a product prototype or client application. Compare those requirements before choosing a route.

IndicStack offers an OpenAI-compatible API for open-weight models, with billing in rupees. Its catalog separates launch models from models available through a requested deployment. Use the catalog to identify candidates, then check their capabilities against representative tasks from your application.

Related terms