Open weights
A model whose trained parameters are publicly downloadable — as opposed to closed-weights models that are only accessible through the maker's API.
An open-weights model publishes its trained parameters (the multi-gigabyte file that is the model). Anyone can download the weights and run them on their own hardware. Meta's Llama family, Mistral's models, DeepSeek's V4 and R1, and several others are open-weights.
Open weights is different from open source. Open source implies you have the code AND the training data AND the freedom to modify and redistribute. Open weights usually means only the trained model file is available; the training data is not published, and the license may restrict commercial use or redistribution.
The practical implications: open weights enable self-hosting (running the model on your own infrastructure), fine-tuning (adapting the model with your own data), and academic research (studying the model's internals). Closed-weights models (Claude, GPT, Gemini) can only be accessed via the maker's API and give you no direct access to the underlying weights.