Safetensors
Safetensors is the format models are originally published in: full-precision weights across many files, built for data-center GPUs and training — not for laptops.
Open a model's page on Hugging Face and the files you see are usually .safetensors — the publisher's original weights at training precision, often split into dozens of multi-gigabyte shards. The name is literal: the format was created as a safe replacement for older pickle-based files, which could execute code when loaded.
These originals are the model's authoritative identity, and the right choice for fine-tuning or for serving on serious GPU hardware. For running locally they are the wrong download: tens to hundreds of gigabytes where the quantized GGUF of the same model is a fraction of the size.
The two formats are stages of one pipeline — publishers release safetensors, the community and vendors convert and quantize them into GGUF for everyone else.
Where you see it on the radar
Model names on the radar link to the original safetensors repo — the model's identity — while the separate GGUF link is the exact file our memory figures were measured from.
See both links on the board →Related terms