I work with a Mac Studio M3 Ultra (512GB) serving DeepSeek V4 Flash on antirez/ds4 ("DwarfStar"). A chat turn took between 6 and 20 seconds. Now it takes 1.6s. Kernels (+21% cold prefill at 64k, bit-exact). DeepSeek V4's sparse attention runs a "lightning indexer" that dominates long-context prefill
