Model guide
GGUF or MLX?
GGUF and MLX are two ways to run a local model. GGUF works across phones and computers. MLX is for Apple silicon. Pick the format that offers the model you want, in a size that fits your memory. The model matters more than the file format, and a cloud model is still stronger on the hardest questions.
What changed. The opening now answers the choice in one pass: GGUF across devices, MLX on Apple silicon, and the model ahead of the format.
Both choices run AI on your device. MLX is built for Apple silicon. GGUF works with runtimes that support a wider range of computers and phones, including Macs. Start with a model that fits your device's memory; the model itself matters more than the label.
What the names mean
GGUF is a model file format. It stores a model's weights and information needed to load them. Loci runs these models through llama.cpp, which supports Apple silicon, other CPUs, and several GPU families. GGUF on a Mac can use the GPU too. GGUF specification, llama.cpp.
MLX is Apple's machine-learning framework. In Loci, “MLX” means a model prepared for the MLX engine on Apple silicon. The framework lets the CPU and GPU work with shared memory. An MLX model is a different download from its GGUF version. Apple's MLX documentation, MLX LM.
| What matters | GGUF | MLX in Loci |
|---|---|---|
| Hardware | A broad range of devices, through a compatible runtime | Apple silicon |
| Model files | GGUF files, sometimes with additional files | Files prepared for the MLX engine |
| GPU use on a Mac | Supported through llama.cpp and Metal | Supported through MLX |
| Availability | Depends on the model and Loci's support | Depends on the model and Loci's support |
| Chat after downloading | Runs locally | Runs locally |
Which should I choose?
If a model already works well for you, you can keep using it. On an Apple silicon Mac, try either version when both are available. Choose GGUF when the model you want is offered in GGUF, or when your device does not support MLX.
Neither choice is always faster or produces better answers. The model, compression, prompt length, available memory, and runtime all affect the result. Compare the same model on your own tasks. A smaller download can save memory, but heavier compression can affect answer quality. Leave room for the conversation and your other apps as well as the model.
What happens when I switch?
The format control changes the models you browse. It does not change the model answering your current chat. Choose Use on a downloaded model when you want to use it.
Each format needs its own compatible download. Switching the list does not convert or delete files, and a download already in progress can continue. You can switch back to find it. Keeping both versions uses storage for both.
Some models are available in only one format. If a model disappears from the list, switch formats or clear your search. A model's name alone does not guarantee that every version has the same capabilities.
FAQ
Is MLX a better version of the model?
MLX describes how the model runs. It does not by itself make the model more capable. Compare the actual model, size, compression, and supported features.
Can I rename a GGUF file to make it an MLX model?
No. The engines need different model files. Download a supported version for the engine you want to use.
Does either format require cloud AI?
No. Both run the model on your device. Downloading models and optional online features such as web search still need an internet connection.
Related guides
See which model fits your device or compare local AI apps for Mac.
Get Loci
Private AI that keeps working.
Download Loci once, then chat without an account, chat upload, or signal.