M3 Ultra + DGX Spark = M5 Ultra-lite?

A Reddit user benchmarked DGX Spark against M3 Ultra in llama.cpp at pp16384, with Spark 1.4× to 3.4× faster across 4 models. Qwen 27B hit 778 t/s vs 340 t/s, while Mistral 128B hit 241 t/s vs 72 t/s. The concrete tuning note is mmap=0: loading fell from minutes to about 20 seconds.

Why it matters: Single Reddit sourcing keeps the score low, but HKR-H/K/R all pass through a concrete local-inference benchmark. The pp16384 setup and 4-model speedups justify featured at the lower edge.