Noema supports five local model formats: GGUF, MLX, ET, CML, and AFM. They do not all install the same way, so the fastest path is to pick the format that matches your hardware and the amount of control you want.
Best default when you want portability, broad compatibility, and the richest tuning options.
Best for fast local inference on Apple Silicon with fewer low-level knobs than GGUF.
Good for Apple-native and mobile-friendly deployments when you want predictable behavior over deep manual tuning.
Apple’s built-in Foundation Model for supported Apple Intelligence devices with the lowest-friction setup.
Best default and most portable.
Pick GGUF when you want the widest hardware coverage and the most tuning. It is the format to reach for when you care about advanced performance controls, sampling precision, quantization choice, and runtime experimentation.
Open Explore, filter or search for a GGUF model, inspect quantization and size, then tap Download.
Best for Apple Silicon speed.
Pick MLX when you want quick local inference on Apple hardware with less backend fiddling. Noema exposes the core settings you are most likely to adjust while the backend stays comparatively opinionated.
Browse Explore for MLX builds when you want Apple-optimized local downloads without the extra GGUF tuning surface.
Best for responsive small on-device models.
Pick ET when you want minimal setup and you are comfortable with fewer knobs. Delegate and backend selection are automatic in the current build, so the experience is intentionally lighter than GGUF or MLX.
Search Explore for ET variants when you want compact mobile-focused deployments with fast startup and fewer runtime choices.
Best for Apple-native execution on Apple accelerators.
Pick CML when you want an Apple-first deployment path for iPhone, iPad, or visionOS and you care more about native execution than about maximum manual tuning. The main format-specific choice is the processing unit used at runtime.
Use Explore to find Core ML builds targeted at Apple devices, then choose the model that matches your supported runtime and storage budget.
Best for the system model on Apple Intelligence devices.
AFM is the lowest-friction option. It is Apple’s built-in system model, so it behaves more like enabling a local capability than downloading a conventional model package.
There is typically no separate weight download in Noema. On supported Apple Intelligence devices, AFM appears as an available local model option once system requirements are satisfied.
Switch to a smaller model, lighter quantization, or shorter context length. GGUF gives you the widest tuning surface if you need to trade quality for speed.
That is expected. These formats intentionally expose fewer runtime controls than GGUF so setup stays predictable on-device.
Verify the device supports Apple Intelligence features and that the built-in system model is available on that OS and hardware combination.
Start with GGUF for the broadest flexibility. Move to MLX when you want Apple Silicon speed, or AFM when you want the simplest Apple-native option.
