Hugging Face / Linux Foundation · 2026-04-08 · major
Safetensors Joins PyTorch Foundation as Vendor-Neutral Standard for Safe Model Serialization
Hugging Face transferred Safetensors' trademark, repo, and governance to the Linux Foundation's PyTorch Foundation. Used by tens of thousands of models on HuggingFace Hub. Roadmap includes PyTorch core integration, direct CUDA/ROCm loading, and FP8/GPTQ/AWQ support.
Safetensors — the format used by most HuggingFace models — becomes a vendor-neutral standard under the Linux Foundation.
What is it?
Safetensors is a file format for storing and loading ML model weights that prevents arbitrary code execution (unlike pickle-based formats). Originally created by Hugging Face, it's the default format for model distribution on HuggingFace Hub and is used by tens of thousands of AI models. On April 8, 2026, Hugging Face transferred its trademark, GitHub repository, and governance to the Linux Foundation's PyTorch Foundation — alongside DeepSpeed, Helion, Ray, vLLM, and PyTorch itself.
How does it work?
The format uses a JSON header (hard-capped at 100MB) describing tensor metadata, followed by raw binary tensor data — no code execution path possible. This enables zero-copy loading by memory-mapping tensors directly from disk, and lazy loading lets you read individual weights without deserializing an entire checkpoint. The roadmap adds: PyTorch core integration (making it the default for torch.save/load), device-aware loading directly to CUDA/ROCm, Tensor Parallel and Pipeline Parallel APIs, and sub-byte quantization types (FP8, GPTQ, AWQ, INT4).
Why does it matter?
By moving to the Linux Foundation, Safetensors becomes infrastructure-grade — no longer tied to a single company's roadmap or business priorities. PyTorch core integration on the roadmap would make it the default serialization system for the world's most-used ML framework. No breaking changes for existing users — all APIs and formats remain the same.
Who is it for?
ML engineers who save, share, or load model checkpoints — effectively the entire open-source ML community