Unfortunately, as I’ve learned recently, it doesn’t look like Deepseek is actually open source.
You can download the model, but unless I’m misunderstanding, that feels comparable to calling Photoshop open source because you can download the .exe file on your computer.
Its MIT licensed. Meaning the code is open but the license is permissible in that copy’s can be subsequently closed. This is unlike with the GPL most generally associated with open source code.
It’s more open-sourced than I thought, but also seems debatable. I don’t know enough about LLMs to properly judge. I would probably stay away from calling it “completely open-sourced” though.
Unfortunately, as I’ve learned recently, it doesn’t look like Deepseek is actually open source.
You can download the model, but unless I’m misunderstanding, that feels comparable to calling Photoshop open source because you can download the .exe file on your computer.
Its MIT licensed. Meaning the code is open but the license is permissible in that copy’s can be subsequently closed. This is unlike with the GPL most generally associated with open source code.
The weights are MIT licensed. The code is, too, but code for these things are uninteresting.
The training data is not open source, and that’s the interesting part of a model.
What’s this? https://github.com/deepseek-ai/DeepSeek-V3
This comment here seems to summarize it well: https://github.com/deepseek-ai/DeepSeek-V3/issues/457#issuecomment-2627016777
It’s more open-sourced than I thought, but also seems debatable. I don’t know enough about LLMs to properly judge. I would probably stay away from calling it “completely open-sourced” though.
You can reweight as you please to whatever dataset you like. They can say what the training data included, but they can’t share the dataset.