3 ms·100 tokens/s, 4-bit Mistral 7B in MLX on M2 Ultra (faster than llama.cpp)3 points by tosh 3y ago