invalid magic characters, expected 'GGUF'

The message

gguf_init_from_reader: invalid magic characters: '<xxxx>', expected 'GGUF'

Parts in <angle brackets> change from run to run.

What it means

Every GGUF file starts with the bytes "GGUF" and this one does not. It is usually a web page, a Git LFS pointer or a safetensors file named .gguf.

Why it happens

• The download saved Hugging Face's HTML page (a /blob/ link instead of /resolve/), or a login page for a gated repo.
• git clone without Git LFS left a small text pointer in place of the weights.
• The file is in another format (safetensors, PyTorch) and needs converting first. A truncated file reports "failed to read magic" instead.

How to fix it

1. Look at the first bytes: head -c 4 model.gguf must print GGUF. A tiny file size is the other giveaway.
2. Download with the Hugging Face CLI: hf download <repo> <file>.gguf --local-dir .
3. For safetensors, convert with llama.cpp's convert_hf_to_gguf.py first.

Guides that cover this

• Quantize Your Own Model to GGUF

Related errors

• llama.cpp: wrong number of tensors