It's open, with a small 'o', and no moral judgements attached to it.

1. You have the weights, so you can run the model yourself on your own hardware. 2. You have the weights, so you can do post-training and shift those weights for your own purposes. It's not the same as training the model, but for many people its fine as what we want is a quantization, or a fine tune, or to create hybrid models.

What they don't give you are the training data, and reproduction instructions but... the toolchain to create the software has never been a part of 'Open Source'.

Even though it feels like a huge loophole, it's technically open source if you deliver the source code, without having a compiler that's available so you force them to recreate the toolchain from scratch.

To me, though, a better analogy is to research science where you'll be happy when they give you the full result set they compiled even if you don't get the raw data which may have IP or privacy concerns, or their often poorly documented lab notes so you can actually reproduce.

What you want to do is run your own experiment, and get your own results... not duplicate theirs directly. Even if reproduction is your aim in science, being unable to reproduce without copious notes sometimes points out that the original experimental process must have been flawed.

LLMs have the same issue. The creation process isn't entirely well documented, and the raw data can't be released since although the company have the right to use certain sources, they can't transfer those rights to others.

Sure, I'm all for opening the whole enchilada. But at least Chinese AI development is more open than the US one. And not just on the weights front, but AFAIK also more open when it comes to the techniques used and lessons learned in the process (DeepSeek at least is). Which is in a sense even more valuable and laudable.

Weights are the cake and source is the recipe. Except you aren’t going to spend millions working through the recipe to arrive at the same exact cake anyway, so why do you care?

The problem with that analogy is that compiling Postgres takes a few minutes. Training a model takes hundreds of thousands of dollars (at least) and specialized hardware.

Having a standardized training set is valuable, though.

Here here! Open weights is the term because it's close to open source but everyone thinks everyone else is an idiot and saying the model is downloadable but you can't recreate it is too difficult for our feeble minds to understand. There are actually open source models out there though, with datasets and training code.