Safetensors Joins PyTorch Foundation Under Linux Foundation Governance — Vendor-Neutral Home for ML's Default Model Format

Hugging Face Blog·high signal

Hugging Face contributed Safetensors to the PyTorch Foundation at PyTorch Conference EU in Paris on April 8, joining DeepSpeed, Helion, Ray, and vLLM as foundation-hosted projects. Safetensors — used by tens of thousands of models as the default format for model distribution on HF Hub — now has vendor-neutral governance under the Linux Foundation, with a roadmap including PyTorch core integration, device-aware loading (CUDA, ROCm), and sub-byte quantization support (FP8, GPTQ, AWQ). This is the most significant ML infrastructure governance move since vLLM joined the foundation.