What is an Open-Weights Model?

Publishing weights lets researchers, companies, and individuals run the model on their own infrastructure, inspect its behavior directly, and fine-tune it for specialized tasks without depending on a vendor's API availability, pricing, or terms of service. Well-known examples include Meta's Llama family, Mistral's models, and Alibaba's Qwen, all of which release weights, usually under a permissive or semi-permissive license, even though the training data and code are often not disclosed alongside them, which is why open weights is a more precise term than open source for these models. Open-weights models matter for data residency and control: an organization with strict requirements about where data can be processed can self-host an open-weights model entirely within its own infrastructure, something impossible with an API-only closed model. The tradeoff is that self-hosting requires real infrastructure investment and ongoing maintenance, and the largest open-weights models still generally trail the top closed frontier models on the hardest reasoning benchmarks.