For the original competition, you can check out https://dawn.cs.stanford.edu/benchmark/#cifar10 . Special note is that these are multi-GPU setups which are much more complicated and out of the reach of most average Joes. Also, they're a pain to engineer around (though necessary at scale, I suppose).

96% is another candidate but much harder, and I think robs the 4-6x speedup in experimentation cycle times that 94% requires.

Yes, sorry I assumed it would be understood that it did the training and maintained quality.

The irony of pointing out missing context and providing insufficient context myself.