2x faster ERNIE-Image/ERNIE-Image turbo 8B on Apple Silicon — MLX inference engine with zero accuracy loss, 46% less memory - View it on GitHub