Safetensors Model Format Explained: Supported Engines and Models

September 27, 2026

What Is Safetensors?

Safetensors is a file format for storing model weights (tensors), developed by Hugging Face. Most models published on Hugging Face ship the weights their publisher trained in this format (.safetensors files). In other words, it is the format that holds the “original" of a model.

Why It Matters

• Loading it does not execute code. The older PyTorch .bin (pickle) format could run arbitrary code when loaded. Safetensors holds only the numbers and a minimal header, so that risk does not exist.
• Fast loading, and you can read only what you need. Files are memory-mapped, so even large models start quickly, and a multi-GPU setup can read just the tensors each GPU needs.
• The common entry point for most tools. Transformers, Diffusers, vLLM and SGLang load it directly, and conversions to GGUF, MLX, AWQ and the rest almost always start from it.

Tips for Running It Locally

• At the original precision (BF16 / FP16) it needs a lot of memory. Plan for roughly two bytes per parameter (about 16 GB for an 8B model). To run a large model on a consumer GPU, look for a quantized build such as GGUF or AWQ.
• Low-precision Safetensors such as FP8 and NVFP4 also exist. They are mainly served with vLLM, SGLang and similar engines on NVIDIA GPUs of a supporting generation.
• Ollama and LM Studio (llama.cpp-based) cannot load it as is. Use a build converted to GGUF.
• Most image and video generation models also use it. For where each file goes in ComfyUI or Diffusers, see the “Files" section of each article.
• MLX, AWQ, GPTQ and EXL repositories also use .safetensors files, but they are listed on their own format pages rather than below.

Sources: Safetensors documentation (Hugging Face) and the huggingface/safetensors README (both as of 2026-09-27).

Our Coverage and Data

Local Model Watch has published 29 article(s) on models available in Safetensors: 29 where the repository itself is in Safetensors, and 0 where we found a Safetensors build of the model. The lists below only include builds we have checked (the publisher’s organization and well-known quantizers); a model missing here may still have a Safetensors build elsewhere. Part of our model format index.

Main Engines That Load This Format

Models Available in Safetensors

“Smallest VRAM tier" is the smallest tier in the requirements table of each article (for other builds, of those builds; for image, video and audio models, always the article’s own table, which counts every component). Leave headroom for context length.

Last updated 2026-09-27 (JST). The explanation at the top of this page was written with the help of AI from the primary sources it cites. The tables under “Our Coverage and Data" are assembled by code from our article log.