Pick from the catalog · Core AI models, typed

A model name is not a product plan. Start with the feature. What goes into the model? What must come out? Is the feature making something new, or describing something that already exists? Those answers usually point to one registry type.

FeatureDescribe the input and output

Typellm, diffusion, or utility

FeatureWhat the person needs

TypeMatch the task

PresetChoose an official registry name

Device pathFollow recipe and measure

A simple chooser

Conversation, drafting, summarising, or code output points to llm. Compare the official Qwen, Gemma, Mistral, Mixtral, GPT-OSS, Phi, Muse Glimmer, and SmolLM2 entries in lesson 02.
A new still image from a prompt points to diffusion. Look at Stable Diffusion or flux2-klein-4b.
A new video from a prompt points to diffusion and the Wan text-to-video entry wan-t2v-1.3b.
A photo label points to utility classification. That is the broad job shared by pvt-v2-b0 and the trained ClothingCNN from the forward course.
Boxes around objects point to utility detection, with yolos-base or yolos-tiny.
A precise object cut-out points to utility segmentation, with efficient-sam-vitt or sam3.
Near-and-far information points to utility depth. A sharper or larger image points to utility super-resolution.
Speech to words points to utility asr. Image/text or audio vectors you can compare point to embedding. Representing or changing text points to encoding.

Device and recipe checks

Once the type fits, look at the exact preset and its recipe. Some LLM families have separate macOS and iOS compression presets. Read that as one chat job with different packing for a Mac or a phone. The registry marks depth-anything-3-small as macOS only. These details can rule out a choice that otherwise looks good.

JobInput and output fit

RegistryOfficial type and name

PackingMac, iPhone, or listed limit

RuntimeProfile the real app path

Go back to the official repository for recipes and the registry. Use the architectures chooser when you need to look again at language versus image shapes. Use the forward ClothingCNN course for a complete trained-model-to-SwiftUI path. Then use From the metal to the model to think from the chip and runtime back up. When the preset is on the device, evaluate it with writing-style eval, generation eval, or classifier eval.

Final rule. The right preset is the one whose typed job, device packing, and recipe fit the feature. The registry narrows the search. Profiling tells you whether the full app path works.

You now have a typed map of the catalog. You do not need to treat every model as a chat model, or every image model as a classifier.

Key concepts

Start with feature input and output, not a model name.
Chat, drafting, summary, or code maps to llm. New still image or video maps to diffusion.
Labels, boxes, cut-outs, depth, and sharpening map to utility vision presets.
Speech to words maps to ASR. Comparable media vectors map to embedding.

Takeaways

You can walk from app feature to type to official preset to recipe.
You can rule out presets by device packing or registry limits before you commit.
You can return to apple/coreai-models, follow the recipe, and profile the real app path.