Was really impressed with Claude ability to port models from cuda PyTorch to MLX, so that’s what I have been doing last three weeks; basically I have Mac to test ported models and Gradient vm with PyTorch and Nvidia gpu, Claude can run code in vm, investigate layers and deep analyze the model mechanics and then reimplement them on local Mac with MLX. So far completed port of various models mainly in audio domain, achieving pretty substantial speeds for inference on my machine. Models ported with numerical parity to originals: facebooks omnillingual ASR, Sam-audio, Nvidia sortformer. Planning on releasing this in repo soon.

I’ve been doing some similar experiments, would love to see the repo once it’s ready.

Local asr/stt on mlx for apple silicon based on Facebook omnilingual model that was just released. Used Claude to port model weights from PyTorch to mlx (that was exciting, claude can do that).

hey, I am preparing for multilingual local Speech2Speech model... Would appreciate if we could connect and maybe help each other in future.