I made a OMR (optical music recognition) model that outperforms the next leading released model (legato) on quartet scans by 6x. It's 28m parameters, with a 2m parameter harness, vs legato which is 100m parameters but with a 900m vision encoder strapped on.