Comparison
GGUFvsSafetensors
GGUF
you downloaded one file with the weights, tokeniser and metadata inside it and pointed the local runner at that.
The single-file model format of the llama.cpp world: weights, tokeniser and metadata together, downloadable at multiple quantisation levels. If you are grabbing a model file for local use, it is almost certainly a GGUF.
Full entry →Safetensors
you download the weight file the Python ecosystem uses and convert it to GGUF, essentially never the other way round.
The standard weight format of the Python and GPU ecosystem (Hugging Face, vLLM, training code): safe to load, memory-mappable, but weights only, without bundled tokeniser. You will convert from safetensors to GGUF, rarely the reverse.
Full entry →