We’re considering bundling a quantized DeepSeek-derived model with an offline desktop application. Customers would receive the model files rather than access a model running on our server.

The download we’re evaluating is a community conversion. Its repository links to a distilled checkpoint, which in turn identifies another model as its base. I don’t want to assume the license on the conversion repository settles the rights for everything in that chain.

For a model redistribution license review, which documents should we collect before sending the package to legal? I’m thinking of the exact checkpoint’s license and model card, any applicable upstream terms, and the converter’s notices—but I’m unsure how to establish which terms apply to the weights versus the conversion code.

Does changing the file format or quantization level introduce anything we should check separately? I’m looking for a practical way to document model provenance tracking, including repository revisions, so we review the actual artifact we intend to distribute rather than a similarly named model.