Safetensors
A format and library for distributing tensor weights without loading Python pickle objects.
Published by Safetensors contributors
How to use it
Use Safetensors when publishing or importing model weights in supported frameworks. The format stores tensor data and metadata, reducing the deserialization risk associated with loading arbitrary pickle objects.
Review accompanying model code, dependencies, and license terms separately. Use a verified source and pin the expected artifact digest when installing model weights in production.
- Access
- Open-source format and library
- License
- Apache-2.0