I'll wait until some more benchmarks are run in this case. Unlike traditional software, vetting a model architecture works better than alternatives is a time and compute intensive process. You really can't just download it and "try it out" outside of general purpose models (which this is not).