GGUF or Safetensors?
Use the format your app and the exact model architecture support. Changing a filename does not convert a model.
Scroll sideways to read all columns. With a keyboard, focus the table and use the left and right arrow keys.
| Format | Where to check support | What to inspect |
|---|---|---|
| GGUF | llama.cpp-based runners and supported Ollama imports | Architecture, quantization, chat template and any extra files |
| Safetensors | The model card and your inference library | Configuration, tokenizer, weight shards and supported architecture |
Choose one exact model release
Search GGUF files, or include the full model name, such as Qwen2.5 GGUF. Open the record and upstream model card. Quantizations are different files: compare size and your runner's supported formats rather than assuming a quantization name is a quality score. A source-only record does not promise a Hugging Bay download.
A filename like model-00001-of-00003.gguf is one part of a split set. Keep all three matching parts together. A vision model may also need a projector; follow its model card and runner instructions.
Check the downloaded bytes
Save the file path, size, publisher and recorded SHA-256 from the same current file row. On macOS, compute a local digest with:
shasum -a 256 /path/to/model.ggufOn Linux, use sha256sum /path/to/model.gguf. Compare the complete digest with a trusted record. A match checks the bytes; it does not prove model quality, permission or successful execution. Then follow the local Llama import example or your runner's instructions.
Sources and checks
Checked 2026-10-26. Install requirements and availability can change; the linked source has current details.
- Hugging Face GGUF format guide
- Import a model into Ollama
- llama.cpp supported model files
- llama.cpp GGUF split and merge implementation
- Hugging Face model cards