What is actually inside your GGUF?

A format label names a recipe, not a type. This reads the header of a GGUF model and shows what it really holds: the types under the label, bits per weight, and the bytes decode reads per token. It runs in your browser. Only the header is read; nothing is uploaded.

A file on your computer Pick a .gguf file. For a split model, select every part.

Or a file on the Hugging Face Hub hf:OWNER/REPO/FILE.gguf, or a huggingface.co link to the file

Try: Qwen2.5 0.5B Q4_K_M Qwen2.5 0.5B Q8_0 Qwen2.5 7B Q4_K_M (split)

Memory bandwidth, GB/s (optional) list every tensor

Loading the Python runtime (a few seconds on the first visit)...

With a bandwidth, the report adds the ceiling: bandwidth divided by the bytes read per token. Real decode lands below it. To measure your own machine's bandwidth and decode speed: pip install llama-roofline, then llama-roofline diagnose. The same report is llama-roofline inspect FILE on the command line, which also reads gated repositories with HF_TOKEN.

This page runs the inspect code of llama-roofline 0.2.3 unchanged, in Pyodide. MIT licensed. Cite as doi:10.5281/zenodo.21842493. The Hub's own file viewer lists tensor types too; what this adds is the on-label share, the bits per weight, the bytes read per token with the embedding excluded and a tied head counted, and the ceiling.