Start with the official source: Apple's official apple/coreai-models repository. It has recipes for working with models and a model_registry.py file that lists named presets. This course uses the typed list from that registry. It does not make up a fourth group or guess model names from a folder.
Apple groups the published presets as llm, diffusion, and utility. These labels help because they tell you the kind of output to expect. An llm writes language. A diffusion model makes pictures or video from a prompt. A utility model does a smaller, focused job such as labels, speech-to-text, vectors, or segmentation (cutting an object out of a picture).
The type is a starting label. It does not mean every preset has the same speed, memory use, or device support. Many LLM families have separate macOS and iOS compression presets. Those are the same chat job packed differently for a Mac or a phone.
Recipes versus a model you trained
In the PyTorch to Core AI course, you train ClothingCNN yourself on Fashion-MNIST. You own that small classifier and how it was trained. The coreai-models repository is different. It gives you recipes and a list of ready-made model presets. You can study them and use them for the jobs they are meant to do.
You may also see folders named vlm and parakeet in the repository. They are extra recipe folders. If they are not in the typed registry presets listed here, do not give them a made-up model name or type. The published list for this course is the typed llm, diffusion, and utility catalog.
The model architectures course explains how language models and image models have different shapes. From the metal to the model reminds you that every chosen shape still becomes work the chip has to run.
Next, look at the largest group in the registry: language models.
Key concepts
- Start at
https://github.com/apple/coreai-modelsfor recipes and the typed list. model_registry.pyis the source of published presets for this course.- The three groups are
llm,diffusion, andutility. - Type tells you the kind of output. It does not promise the same speed or device support.
vlmandparakeetfolders can appear in the tree. They are not typed catalog presets here.
Takeaways
- You can name the official repo as the source of recipes and registry types.
- You can describe
llm,diffusion, andutilityby output shape. - You can choose input and output first, then type, then a named preset.