I’m evaluating DeepSeek models for a product prototype and trying to understand what “open” means for a real deployment. I don’t want to assume that the terms are identical across model versions or that using hosted inference has the same implications as downloading weights.

When checking the DeepSeek model license, which version-specific documents or restrictions should a team review before planning commercial deployment terms? I’m interested in practical experiences, especially where the model is modified, fine-tuned, or bundled into a paid service.