Comment by antirez

The README lacks the most important thing: how many more tokens/sec at the same quantization, compared to llama.cpp / MLX? It is worth to switch default platforms only if there is a major improvement.