This model was converted to MLX from deepseek-ai/DeepSeek-R1-Distill-Qwen-32B, using mixed 2/6 bit quantization. This scheme preserves quality much more than a standard 2-bit quantization.