I finally got round to finishing a much longer post on what we can conclude about Watt-hours per query based on the InferenceMAX benchmarks, and the limitations with this muxup.com/2026q1/per-q...

Congratulations!

I'd looked at Apertus and it seems they couldn't have a similar ratio based on the publicly announced GPU hours awarded to the project (10M per the Sep press release) and the documented GPU hours on training (at least 6M per the paper). But that's perhaps reflected in the results.

Thanks for confirming. The fact experiments will use substantial compute was clear, but I hadn't realised it would exceed the pretraining cost (and hadn't seen it disclosed in other write-ups). Great to have this level of clarity and transparency.

For slide 15 on the eval talk: I understand pretraining for the two base models was ~1.24M GPU hours. Is this really saying >1.4e6 (1.4M) GPU hours for post training or am I interpreting it incorrectly?

Me too, I even like to imagine it extends to leisure centres outside of Wales!

Here's something to put on in the background while you draft the submission to propose it as a new SI unit www.youtube.com/watch?v=f6eP...

"FWIW, I think that people overindex on what “superset” means. C++ is considered a superset of C even though it isn’t literally source compatible. I believe that Mojo will end up with the same relationship to Python in the fullness of time."