Definition
An open-weight model is a trained model whose learned parameters, or weights, are available for others to obtain and run. A release usually includes the configuration and other files needed for inference, such as a tokenizer for a language model.
The accompanying license defines what users may do with those files. It may permit local inference, modification, fine-tuning, or redistribution, but the term “open weight” does not promise all of these rights by itself.
Simple example
A model provider publishes a downloadable checkpoint, its architecture configuration, a tokenizer, and instructions for running inference. A team downloads the files and serves the model on its own hardware instead of calling the provider’s API.
The provider does not publish the training dataset or the code used to prepare the data and train the model. The release is still open weight because the trained parameters are available to run. That alone does not make the model open source.
Why it matters
Access to weights gives engineers deployment choices that an API-only model cannot offer. They can run inference in their own environment, measure performance on specific hardware, and adapt the model when the license allows it. Self-hosting can keep inputs within a controlled environment and continue working where external connectivity is unreliable.
That control comes with operational work. The team must provide suitable hardware. It must also operate and update the serving stack. Application-level evaluation remains necessary regardless of the hosting model.
One important nuance
Open weight and open source are not synonyms. Under the Open Source AI Definition, open source AI also requires training and inference code, sufficiently detailed information about the training data, and terms that permit use, study, modification, and sharing.
“Open weight” is not one standardized license category. Two releases carrying that label may grant different rights or impose different conditions. Check the actual license before using, modifying, or redistributing a model.