I assume those don't just work automatically with an off-the-shelf gguf. What do you need in your local inference stack to take advantage of M5's neural accelerators?