Let's stop saying that open weight models are open source. They are just a software that you can run on your own hardware (if you have a ton of money to spend to get it anyway) but they are not open, they are as open as any proprietary software that you can run on your PC is.

You can't study them, you can't modify them, you can't improve them, you can't look into how they are built etc.

I never said they were open source. Everything else you wrote is factually incorrect.

When you run a proprietary program (like Adobe Photoshop or a closed-source game) on your PC, you get a compiled binary of machine code. The high-level logic is stripped out, variables are obfuscated, and the EULA legally forbids you from decompiling, reverse-engineering, or modifying the binary.

With an open model, you get the actual Python/C++/CUDA code defining the exact mathematical structure of the model. And you get a weights file, which is an unencrypted, open array of floating-point numbers representing every synaptic strength in the network.

So you can inspect the exact mathematical state of the model at any arbitrary moment during inference -- in other words, you can study open models. You can fine-tune or quantize them -- in other words, you can improve them quite significantly! And much much more.