Audio version — Estimated duration: 5 min 18 sec

Model Conversion — Hugging Face → GGUF

Most models on Hugging Face are distributed in the raw PyTorch format (a folder of .safetensors or .bin files). llama.cpp cannot use these directly — they must first be converted to GGUF.

In most cases, you don’t need to do this. Pre-converted GGUF files are available from community members like Bartowski, MaziyarPanahi, and others. Only follow this guide if:

  • The model you want does not have a GGUF version available.
  • You want to convert your own fine-tuned model.
  • You want a custom quantization not available pre-built.

How It Works

Raw model (safetensors) → convert.py → FP16 GGUF → quantize.exe → Q4_K_M GGUF

Step-by-Step

1. Install Python Dependencies

pip install torch transformers sentencepiece

2. Download the Raw Model from Hugging Face

git lfs install
git clone https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3

GitHub LFS is required for large model files. Install from git-lfs.com.

3. Convert to FP16 GGUF

From the llama.cpp directory:

python convert.py "C:\AI\raw-models\Mistral-7B-Instruct-v0.3" ^
  --outfile "C:\AI\models\Mistral-7B-Instruct-v0.3.fp16.gguf" ^
  --outtype f16

4. Quantize to a Smaller Format

quantize.exe "C:\AI\models\Mistral-7B-Instruct-v0.3.fp16.gguf" ^
  "C:\AI\models\Mistral-7B-Instruct-v0.3.Q4_K_M.gguf" ^
  Q4_K_M

5. Verify

./bin/Release/llama-cli.exe -m "C:\AI\models\Mistral-7B-Instruct-v0.3.Q4_K_M.gguf" -p "Hello!"

Supported Quantization Types

Pass the quant type as the last argument to quantize:

TypeSize vs FP16Quality
Q2_K~17%Poor
Q3_K_M~25%Low
Q4_0~27%Medium
Q4_K_M~29%Good (sweet spot)
Q5_K_M~34%Very good
Q6_K~40%Excellent
Q8_0~50%Almost lossless

Converting Without a GPU

The convert.py and quantize steps are CPU-only — they don’t need a GPU. The FP16 file will be large (2× the final size), so ensure you have enough disk space.

One-Line Conversion (convert-hf-to-gguf)

Newer versions provide a convenience script:

python convert-hf-to-gguf.py "C:\AI\raw-models\Mistral-7B" ^
  --outfile "C:\AI\models\model.gguf" ^
  --outtype q4_k_m

Last Updated: 2026-07-19