Hugging Face Transfers Safetensors to PyTorch Foundation

Hugging Face announced it is transferring Safetensors, its secure tensor serialization format, to the PyTorch Foundation, placing one of the machine learning ecosystem’s most widely used file formats under open governance.

The move hands stewardship of the format to the Linux Foundation-hosted PyTorch Foundation, which already oversees the PyTorch deep learning framework and a growing portfolio of related open-source projects.

Safetensors was developed as a secure alternative to Python’s pickle-based serialization, which has long been flagged as a security risk due to its ability to execute arbitrary code during deserialization. The format stores model weights in a simple, memory-mappable binary layout that eliminates that attack surface while offering faster load times.

Since its release, Safetensors has become a de facto standard across the ML ecosystem. Major model repositories on the Hugging Face Hub default to the format, and it has seen adoption across frameworks and tools beyond PyTorch, including libraries for model inference and deployment.

The transfer to the PyTorch Foundation follows a pattern of major AI infrastructure projects moving to neutral governance structures. The foundation, which counts Meta, Google, Microsoft, Amazon, AMD and other major technology companies among its members, provides a vendor-neutral home for projects critical to the deep learning stack.

For Hugging Face, the decision reflects a strategic choice to cede direct control of a project it created in exchange for broader industry buy-in and long-term sustainability. The company will continue to contribute to Safetensors development but will no longer be its sole steward.

The PyTorch Foundation said the addition aligns with its mission to support open-source projects that advance the AI ecosystem. Safetensors joins PyTorch itself and other foundation-governed projects including ExecuTorch and TorchServe.

The transfer is expected to be completed in the coming weeks, with governance details to be finalized by the foundation’s governing board.