Depends on the release. DeepSeek has used different license terms across releases, and the pattern to watch is labs that write their own licenses instead of Apache 2.0 or MIT. - Some releases permit commercial use with attribution. Others carry specific model license agreements. Read yours. - Check the weights-vs-outputs split: some licenses allow commercial use of outputs while restricting redistribution of weights. For API products, that distinction matters. - Research-only licenses are the quiet trap. The word "research" in a license title is doing a lot of work. Respect it. - Treat every new checkpoint as a new licensing event. Do not rely on a tweet. The test is simple: does your exact release grant commercial use? Not the family. Not last year's release. Yours. PrivateLLM deploy sets up the licensed model on your own AWS for $50 plus usage.