| Format | What it is | Opens in |
|---|---|---|
| GGUF | One file, already quantized, with the settings inside. The most widely supported format for local use. Very large models are sometimes split into several part files. | llama.cpp, Ollama, LM Studio, Jan, Lemonade, Docker Model Runner, KoboldCpp, llamafile, GPT4All, Msty, and PocketPal AI and ChatterUI on phones |
| MLX | Apple’s array framework. Quantized weights laid out for Apple silicon, often faster than GGUF there. | LM Studio, MLX LM, Msty, Inferencer, exo and others, on a Mac only |
| Safetensors | The original weights, usually 16-bit, plus quantized variants such as AWQ, GPTQ and FP8. Built for servers. | vLLM, SGLang, Hugging Face Transformers, Text Generation WebUI |
Rules of thumb
- Windows or Linux with a graphics card, one user: GGUF in LM Studio or Ollama.
- Mac: try the MLX build first in LM Studio; use GGUF if a model has no MLX version.
- Phone, tablet or other mobile device: one whole GGUF file in PocketPal AI. Mobile apps can’t open a model that is split into parts (mobile devices).
- A server for a team or an app: safetensors in vLLM, which batches many requests.
You rarely have to choose by hand: every model page lists the files that exist and picks the best app for your hardware. Browse all apps to see which formats each one reads.