Nanbeige4.3-33B Q4 on a MacBook Pro M5 Max, mlx-lm vs llama.cpp vs Kapitsa, Prefill/Decode 512 to 32K range.
. hey there, I measured Nanbeige4.3-33B (dense, 4-bit) on three Apple silicon engines: mlx-lm, llama.cpp and Kapitsa. The comparison covers prefill, decode, time to first token, follow-up latency and memory, at prompt lengths from 512 to 32K tokens.