I reviewed the AlphaFold3 paper from DeepMind for the journal Nature. I tried really hard to get the editors to demand that DeepMind release the code (even an executable) so people could do the many high-throughput studies we saw for AF2 (see image from my review). I failed. So just a server 4 now.

How many residues could it do? We have tested some with Google cloud and they were able to get 6000 and they claimed 9000 on v2. This can’t be achieved on any NVIDIA GPU with v2 including gracehoppers and SXM. They claimed it can’t parallelise yet they were doing something to achieve those sizes.

And some serious restrictions! Thanks for fighting the good of humanity though.

Thanks for this effort! Do you think they'll ever release?