Choose a model that fits
Safetensors and GGUF
Recognize two common model formats and choose between training-oriented tensor storage and portable local inference packages.
safetensors and GGUF solve related but different problems.
Safetensors stores tensors as data without using Python pickle. It is designed for safe and fast loading, including partial and zero-copy access. Model repositories often split large weights across several .safetensors files.
GGUF is a single-file format designed for GGML-based inference runtimes. It stores tensors plus standardized metadata. It also supports many quantization types.
You will often see the same model published both ways:
model-00001-of-00004.safetensors
model-00002-of-00004.safetensors
...
model-Q4_K_M.ggufUse the format your runtime expects. llama.cpp loads GGUF. Ollama can manage compatible models for you. Transformers commonly loads Safetensors repositories. MLX models use artifacts prepared for Apple’s MLX ecosystem.
A conversion is another build step. Record the original model, converter version, quantization method, output hash, and test results when reproducibility matters.
Lesson completed