Converted locally from alphaZimuth/Hy-MT2-30B-A3B-APEX-GGUF/Hy-MT2-30B-A3B-APEX-Imatrix-I-Quality.gguf for oMLX/MLX inference.

The source GGUF mixed precision policy is preserved by mapping IQ4_XS/Q5_K/Q6_K/Q8_0 tensors to MLX affine 4/5/6/8-bit modules with group size 64. Quantization parameters and ordinary floating weights use BF16; expert routing biases remain F32.

Source tensor types: {'F32': 287, 'IQ4_XS': 60, 'Q5_K': 30, 'Q6_K': 272, 'Q8_0': 117}.