Definition
A restricted-weight model makes its trained parameters available under conditions that limit access or permitted use. A provider might approve recipients individually, allow only research use, or require separate permission for commercial deployment.
Recipients obtain the weights, but that alone does not guarantee local inference. Execution may also require supporting artifacts and compatible software and hardware. The release terms must permit running the model, and they determine whether recipients may fine-tune it or redistribute its weights.
With API-only access, users send requests to a hosted service without receiving the weights. Restricted-weight access gives eligible users the weights, but places conditions on obtaining or using them.
Simple example
Suppose a provider releases a language model to approved researchers. Its terms allow local inference and fine-tuning for research, but prohibit commercial use and sharing the checkpoint outside the approved team.
A university team downloads the weights and evaluates the model on its own GPU. The same permission would not cover deploying it in a paid customer-support service or publishing the fine-tuned checkpoint.
Why it matters
Weight access can make a deployment technically possible while the release terms rule it out. Check permissions for the intended workload before investing in serving infrastructure or model adaptation. Restrictions on redistribution can also affect whether you may package weights in an application or pass a modified checkpoint to another team.
Use restricted-weight access when its terms permit your workload and you need the parameters directly, for example to evaluate or adapt the model. Avoid it when your required use or distribution falls outside those permissions.
One important nuance
“Open weight” and “restricted weight” are descriptive labels, not standardized license categories. Some classifications include weights with usage restrictions under open weight. Others put them in a separate restricted-weight category. Check how a source uses the labels. Neither label guarantees that the model is open source.
Download gating and usage restrictions are separate. An access gate controls who can obtain the weights and may require acceptance of terms. The applicable license or agreement defines what recipients may do with them. A freely downloadable checkpoint can still carry usage restrictions.