safetensors/safetensors

Health v2 · maintenance only

Adoption not part of the score

3876 stars · 354 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

Safetensors is a simple, secure file format and library for storing and distributing tensors, designed as a fast zero-copy alternative to pickle-based weight storage. It provides bindings for Python (PyTorch, TensorFlow, JAX, NumPy) and a native Rust implementation.

Use cases

• save model weights without pickle security risks
• load pytorch model checkpoints fast with zero-copy mmap
• convert pickle .bin weights to safetensors format
• read tensor metadata like shapes and dtypes via HTTP range requests
• share model weights safely on the hugging face hub
• load only part of a large tensor for multi-GPU sharding

When to choose

• you need to store or distribute neural network weights securely
• you want faster checkpoint loading than pickle-based .bin files
• you work in the Hugging Face ecosystem with Transformers or Diffusers
• you need lazy or partial loading of large tensors

When to avoid

• you need to serialize arbitrary Python objects or non-tensor data
• you require a format with built-in compression
• you need full model architecture serialization rather than just tensors

Facets

library · maturity stable

serialization machine-learning deep-learning developer-tools machine-learning deep-learning large-language-models developer-tools files cross-platform python rust cli tensors model-weights zero-copy pickle-alternative hugging-face file-format pytorch tensorflow jax mmap

Member repositories

For agents

markdown · JSON · MCP: product_card(name="safetensors/safetensors")