Splash is fast, 2x faster prefill and decode than similar quants on oMLX. At least in my limited, and unscientific, 36 hour+ tests

https://github.com/JamesFranc/refactored-parakeet/blob/main/...