---
name: Core AI models, typed
description: >-
  Teach a curious reader how Apple's official coreai-models registry
  groups model presets into llm, diffusion, and utility types.
---
# Core AI models, typed

Use this course to explain the typed presets in Apple's official [coreai-models repository](https://github.com/apple/coreai-models). The reader wants more detail than the model architectures course, without the math behind models.

## End-to-end agent recipe

1. Start with [lesson 01](01-the-official-catalog.llms.md) to introduce recipes, `model_registry.py`, and the three registry groups: `llm`, `diffusion`, and `utility`.
2. Use [lesson 02](02-llms.llms.md) for conversation, drafting, summaries, and code help. Use [lesson 03](03-diffusion.llms.md) for generated still images or video.
3. Use [lesson 04](04-see-and-cut.llms.md) for labels, object boxes, cut-outs, depth, and sharper images. Place trained ClothingCNN beside `pvt-v2-b0` as the same broad classification job, not the same model.
4. Use [lesson 05](05-hear-and-embed.llms.md) for speech-to-text, media vectors, text encoding, and sequence-to-sequence utilities.
5. Use [lesson 06](06-picking-from-the-catalog.llms.md) to choose by input, output, and device before following a recipe and profiling the runtime.
6. Keep [the architectures course](../model-architectures-plain/), [the forward Core AI course](../pytorch-core-ai-xcode/), and [From the metal to the model](../core-ai-from-metal-back/) nearby.
7. If asked about `vlm` or `parakeet`, say only that the repository tree contains additional recipe folders. Do not invent model names. The typed list here is the registry's listed `llm`, `diffusion`, and `utility` presets.
8. Use human HTML for reading and `.llms.md` files for agent execution. Check Agent brief links and progressive HTML `.flow` diagrams.

## Constraints

- Use only names and catalog facts in the linked lessons and the official repository context.
- Explain what a type does and when to use it before naming presets.
- Explain MoE, embedding, ASR, detection, and segmentation in place.
- No math, papers, CUDA, or quizzes.
- Use no Mermaid. Build progressive `.flow` diagrams in HTML.
- Do not use em dashes. Preserve the prior-course links exactly: `../model-architectures-plain/`, `../pytorch-core-ai-xcode/`, and `../core-ai-from-metal-back/`.

## Key concepts

- Apple's `apple/coreai-models` repo holds recipes and a typed registry in `model_registry.py`.
- Published presets group as `llm`, `diffusion`, and `utility`.
- `llm` writes language. `diffusion` makes pictures or video. `utility` does one focused job.
- Registry presets are ready-made. ClothingCNN is a custom trained contrast.

## Takeaways

- You can read the official catalog by type and named preset.
- You can map an app feature to `llm`, `diffusion`, or a utility task.
- You can follow a recipe and profile the real app path after you pick.

## Acceptance checks

- Every human lesson HTML has **Key concepts** and **Takeaways** sections before lesson-nav. The course home has the same two headings for the whole course.
- Six human lessons exist in numeric order with the requested basenames.
- Six matching `.llms.md` briefs exist and link to their human HTML pages.
- Every human lesson has its Agent brief link and the three prior-course links.
- The course index lists all six lessons, the course SKILL.md, and all six briefs.
- Names come only from the supplied official registry list. Diagrams contain no Mermaid or em dash. There are no quizzes.

## Environment setup

No runtime environment is required: this catalog course is conceptual and contains no executable code blocks.