That was a joke in the release video. The Pythia model is already released at [1] and the deltas for the LLaMa model should be up here [2] in the next few days.

[1] https://huggingface.co/OpenAssistant/oasst-sft-4-pythia-12b-...

[2] https://huggingface.co/OpenAssistant/oasst-llama-based-model...

Unfortunately [2] is just a placeholder for now, but it does look like the intent is to publish the weights.

It's also relatively cheap to make your own llama-30 weights, the real value of OpenAssistant is in the training data, and all of that data has been made available.

The OpenAssistant effort gets an A+ for open source contributions.

The announce video by Yannic contained a (lengthy) gag to that effect, has it been taken out of context or did now something actually happen?

https://youtube.com/watch?v=ddG2fM9i4Kk&t=132

It's easy to miss but after the negative build-up he says: "and... I'm kidding!"

Dangerous gag, he said “I’m joking” so quickly it’s very easy to miss. I imagine the commenter is not alone in having that wrong impression.

Oh, ha, yeah this is exactly the gag I fell for. I just noped out of the video and wrote off the project as this was the first I ever heard of them, and their website just has a signup and no downloads I could see.

Too bad my original comment is too old to edit.

Unless something changed, I thought it was that they literally cannot legally release the weights that are based on LLaMA (except maybe with an xor thing) so they’re going to train it based on something else

Is any of the Open Assistant stuff based on LLaMA? I thought they release (at least some version) before LLaMA even dropped?

Yes, there’s also something based on Pythia but it’s a smaller model

IIRC, the video said they will train it on a properly open-source model as well.

There was a dumb joke along those lines in an announcement video, meant as a jab at OpenAI. It's easy to miss the "just kidding". (I did, initially.)