“Many open-source models are trained using distillation, where a smaller model is trained to emulate the outputs of a larger, more capable proprietary model.”