← All courses ← Course home

Lesson 02 of 06

LLM: chat and code assistants

The llm type is the catalog's path for writing language. Use it when the app needs a reply, a draft, a summary, or a code-shaped answer.

Agent brief (llms.md)

An LLM (a large language model) reads some text and continues it. In an app, that can look like a chat reply, a rewrite, a summary, or help with code. The model is not a photo labeler just because the text talks about a photo. Start with the output. If the output is useful language, look at the llm presets.

InputInstructions and context
llmRead context and continue
InputPrompt and conversation
ModelLanguage preset
OutputReply, draft, summary, or code

Qwen families

The registry includes qwen2.5-1.5b-instruct, plus an 8bit-kv variant. It also includes Qwen3 at qwen3-0.6b, qwen3-1.7b, qwen3-4b plus an 8bit-kv variant, and qwen3-8b. For code work, the catalog includes qwen3-coder-30b-a3b-instruct. The sizes are a simple trade: a larger model can hold more patterns, and a smaller one is easier to fit and run.

Gemma families

The Gemma entries are gemma3-4b-it and gemma3-12b-it. The Gemma 3n entries are gemma-3n-e2b, gemma-3n-e4b, and their it variants. The family names help you group related presets. When you follow a recipe, you still need the exact named preset.

Mistral, Mixtral, and GPT-OSS

mistral-7b-instruct-v0.3 appears with macOS and iOS presets. That is a helpful reminder that one chat job can be packed differently for a Mac or a phone. mixtral-8x7b-instruct-v0.1 uses a mixture-of-experts style (MoE). Different parts of the job can go to different expert groups. The registry also includes gpt-oss-20b.

Phi, Muse, and SmolLM2

The Phi entries are phi-4-mini-instruct, phi-3-mini-instruct, and phi-3.5-mini-instruct. The catalog includes muse-glimmer-30b with a drafter assistant. It also includes smollm2-1.7b-instruct, smollm2-360m-instruct, and smollm2-135m-instruct.

One family, several packing choices

Many families have separate macOS and iOS compression presets. In simple words, the same kind of chat job has been packed differently for a phone or a Mac. The 8bit-kv names are another packing choice you can see in the Qwen entries. Do not treat a device preset as a new kind of chat job.

NeedChat or code output
FamilyCompare catalog families
DeviceMac or phone packing
RecipePrepare, run, and measure

The language-model lesson explains the words-to-next-token shape. The forward course shows the ClothingCNN path, which is a different kind of job. From the metal to the model shows why this language work still ends up on the chip.

Use this type when the app must write language. Pick among the official names only after you know the device, the context, and the response you need.

When you want pictures or video instead of words, move on to diffusion.

Key concepts

  • llm reads instructions and context and writes language.
  • Named families include Qwen, Gemma, Mistral, Mixtral, GPT-OSS, Phi, Muse Glimmer, and SmolLM2.
  • Many families have separate macOS and iOS compression presets.
  • 8bit-kv is another packing choice, not a different job.

Takeaways

  • You can pick llm when the app must write language, not label a photo.
  • You can name official presets from the registry, not invented model names.
  • You can treat macOS versus iOS packing as the same chat job packed for a Mac or a phone.