About mlxradar

mlxradar lists the models you can run on Apple silicon in two formats: MLX, Apple’s open-source machine learning framework, and Core AI, Apple’s on-device model format. It brings both catalogues into one list so you can see what exists, how big it is and whether it fits your Mac.

Where the data comes from

MLX builds come from mlx-community on Hugging Face, where MLX conversions are shared. The full listing is read every three hours: 5,842 builds at the last refresh. Each build’s base model, task, quantisation, parameter count, licence and download count come from Hugging Face.

Core AI models come from Apple’s apple/coreai-models repository, read every hour. Apple publishes export recipes there rather than finished packages: each entry names an open model, the platforms it targets and the compression, precision and context length Apple uses. The date shown for each entry is the first commit to Apple’s registry that included it.

Builds and recipes for the same upstream model are joined into one page, so 2,357 models are listed, 34 of them in both formats.

What is estimated

Weights sizes for MLX builds are worked out from Hugging Face’s parameter counts and each build’s quantisation, as Hugging Face does not list them directly. For Core AI, sizes are worked out from the parameter count at Apple’s compression setting.

Memory needed is the weights, plus a fifth for context and working memory, plus 2 GB for macOS and your apps, rounded up to a whole gigabyte. On the “will it fit” meter, green means the Mac has at least 60% more memory than that; amber means it should run if you close other apps. Long contexts and large batches need more.

mlxradar does not run benchmarks, so it does not show speeds.

What the labels mean

New marks a model whose first MLX build or first Core AI entry appeared in the last 14 days. Trending marks the two dozen models with the highest Hugging Face trending scores across their MLX builds.

Independent

mlxradar is not affiliated with Apple Inc. or Hugging Face. It is made by bzwrd, which also runs aimodelradar.app.