Hugging Face hands Safetensors to the PyTorch Foundation for neutral governance

On April 8, 2026, Hugging Face announced that it is moving the Safetensors weight-storage format under the PyTorch Foundation. For the open-source ML ecosystem this is not a change of technology but a change of governance: the format, the API and compatibility stay the same, but the project's development now belongs to a neutral body under the Linux Foundation.

Why Safetensors exists

Safetensors emerged as a response to a problem the ML community tolerated for a long time: many model weights were distributed in pickle-based formats, which means they could execute arbitrary code on load. While model sharing was a niche practice, the risk was often ignored. But as open repositories grew and models were reused at scale, that approach became too dangerous. Hugging Face bet on a simple format that stores metadata in a JSON header with a hard size limit and keeps it separate from the raw tensor data.

Technically, the format solves several practical problems at once. It supports zero-copy loading, meaning weights can be mapped straight from disk without extra copying, as well as lazy loading, where you read only the parts of a checkpoint you need instead of deserializing the entire file. According to Hugging Face, this balance of safety and performance helped Safetensors become the default format for distributing models on the Hugging Face Hub and beyond. Today it is used in tens of thousands of models of all types, from language to multimodal.

What changes now

Safetensors is now a foundation-hosted project within the PyTorch Foundation, which operates under the Linux Foundation. For the project, this means a move to vendor-neutral governance: the trademark, the repository and the formal governance rules are no longer tied to a single company. Day-to-day work is not reset, though: the two key Hugging Face maintainers, Luc and Daniel, remain in technical leadership and continue to run the project.

"For most users, nothing changes."

That is perhaps the main signal for the market. Developers do not need to migrate to another format, rewrite code or wait for breaking changes. Hugging Face stresses separately that existing Safetensors files, current APIs and Hub integration work as before.

For the community, however, the path to becoming a maintainer is being formalized: governance rules and the list of maintainers are published in open documents, and companies building products on top of the format get a more durable institutional footing.

What's next for the project

The most important next step is closer integration with PyTorch. Hugging Face says it is working with the framework team so that Safetensors can be used inside PyTorch core as the serialization system for torch models. If this reaches production status, the format will strengthen its position not only as a safe alternative for hubs and repositories, but also as a basic mechanism for exchanging weights within the PyTorch ecosystem itself.

Over the coming months, the team plans to pursue several directions at once:

• device-aware loading and saving, so tensors load straight onto CUDA, ROCm and other accelerators
• APIs for Tensor Parallel and Pipeline Parallel, where each rank or stage receives only the weights it needs
• formal FP8 support
• support for block-quantized formats, including GPTQ and AWQ
• support for sub-byte integer types

The context in which these features will develop matters as much as the list itself. Inside the PyTorch Foundation, Safetensors can tackle these tasks together with other infrastructure projects rather than in parallel to them, alongside PyTorch, DeepSpeed, vLLM, Ray and Helion. For developers, that raises the odds that the format evolves as a common ecosystem standard rather than as a useful but external tool owned by one company.

What this means

Safetensors' move to the PyTorch Foundation formalizes what had already happened de facto: the format grew from an internal Hugging Face initiative into shared open-source ML infrastructure. If neutral governance does speed up support for new data types, parallel loading and integration with PyTorch core, the market will get not just a safer weight format but one of the basic building blocks for the next generation of ML tooling.