Well, you can download and run R1 locally, where you could presumably benchmark it.

But what computational power went into the model they are working with?

@briannawu @bouncing it doesn't matter. According to openAI it's trained on the output of their model. Meaning it's a derivative model, most of the hard work was already done. based on a few headlines I scrolled passed.

I think the only evidence it must have been less computing power is (1) their supposedly lower budget, (2) the fact that they aren’t allowed to buy nice NVIDIA chips.

So yeah there’s some “trust us bro” in there for sure.

I literally just watched this. The presenter is a long-time Pi-et-al reviewer. (I have not tried any of it myself, but I used to know him years ago and I trust him to be honest about his findings.)

https://www.youtube.com/watch?v=o1sN1lB76EA

This may be helpful:

https://medium.com/@isaakmwangi2018/a-simple-guide-to-deepseek-r1-architecture-training-local-deployment-and-hardware-requirements-300c87991126

You have to differentiate between training costs and running costs. Running costs can be independently verified since the model is open source, training costs cannot.

oh you're here! (kinda) that's hilarious ~Chara