Hugging Face Forms Dedicated PyTorch/MPS Team for Apple Silicon
Hugging Face has formed a dedicated engineering team focused on PyTorch Metal Performance Shaders (MPS) support for Apple Silicon (M1–M4). Initial deliverables already shipped: torch.sort and torch.multinomial implemented as native MPS shaders, with a 5× improvement in safetensors loading speed on MPS. Next milestone: flex attention on MPS. The team's stated performance target is 100× improvement over baseline Apple Silicon PyTorch performance. This is the first organized institutional commitment to serious Apple Silicon performance work within the HuggingFace engineering org.
Why It Matters
Apple Silicon (MacBook Pro, Mac Studio, Mac Pro) is the dominant development hardware for the AI builder community. Serious MPS performance work directly accelerates local model development and evaluation cycles. Combined with llama.cpp's 100K star milestone and Qwen3.6 27B reaching near-Opus quality on MacBook Pro, HuggingFace's team formation signals a coordinated ecosystem bet that Apple Silicon local AI will be a major deployment target within 12 months.