Are you running the Moonshine model via ONNX/CoreML or native ggml/mlx bindings? how is the first token latency and memory footprint compared against Whisper small.en on Apple Silicon