Wait, are open weights effectively open source?

It annoys me when people call open weight models open source. You get a directory of safetensors files and a license, with no training data, no data pipeline, no training code, and no hyperparameters, so you can’t rebuild the model from what was published. The G7 recently proposed a four-tier openness spectrum, and Stephen O’Grady checked 40 open models against it and found that zero of them qualify as open source AI.

Then I thought about distillation and mostly talked myself out of the complaint.

You can make new models out of them

Making derivative works is the freedom that actually matters, and open weights give you that completely. You can fine-tune, merge, prune, or distill, running the teacher on your own GPUs to generate training data for a new model that belongs to you. Qwen’s Apache 2.0 and DeepSeek’s MIT licenses explicitly allow it.

This is how frontier models get built now. Gemma, Qwen3, DeepSeek-V4, and Nvidia’s Nemotron 3 Ultra all distill from teachers, and Qwen3’s report puts it at roughly 1/10 the GPU hours of RL with better results. DeepSeek shipped the R1-Distill family by pouring R1’s reasoning traces into Qwen and Llama students, and a recent paper on on-policy distillation finds the student can pass the teacher when the teacher has genuinely new capabilities to offer.

The weights might be better than the source

DeepSeek-V3’s pretraining run took 2.7 million GPU hours. If they’d published the data and the training code instead of the weights, almost nobody could have done anything with it. What they did publish is the compressed result of all that compute, small enough to fit on a USB drive and cheap enough to modify on rented GPUs, so you get to skip the expensive part.

Provenance is the real loss. You can’t see what went into the model, which matters for security and trust, and licenses vary more than people assume since derivatives inherit them (a distill of Llama is still stuck with the Llama community license). The best objection to my reasoning here is that me being unable to use the training data doesn’t mean nobody can, and a research group with real compute would get a lot out of having it.

But if the test is whether a stranger can take your artifact and build a better one without asking you, open weight models pass. The label still bugs me and I have to admit it’s mostly a semantic complaint now.