Open Weights Glossary
Open weights are trained model parameters released so that developers and operators can download them, run them, and build systems around them without sending every request back to the lab that trained the model.
Open weights are not automatically open source. The training data may remain closed. The training code may remain closed. The license may restrict commercial use, military use, fine-tuning, redistribution, or certain deployments. But open weights still matter because they let the model leave the provider’s server.
That is the sovereignty point. A closed API gives access to capability but keeps the operator inside the vendor’s pricing, policy, uptime, and product decisions. Open weights let the operator run the model locally, in a private data center, on a university cluster, or inside an institutional environment governed by the operator’s own rules.
The irony, as of 2026, is that some of the most practically sovereignty-supporting open-weight models have come from Chinese labs — especially Qwen and DeepSeek — while the leading U.S. frontier labs have kept their strongest systems closed. The old map that placed “openness” in the liberal West and “control” in China is no longer good enough.