# apfel on a Mac: inspect a local model through prompts, chat and an API

Follow the recorded command-line routes, verify compatibility and distinguish model-local processing from the rest of your workflow.

[HTML](https://lukejbyrne.com/library/guides/x87fiwoxifq/)
Content language: en
Author: Luke Byrne
Video published: 2026-07-09
Guide published: 2026-10-31
Sponsored coverage: Scrimba

[Watch the video](https://www.youtube.com/watch?v=x87FIWoXIfQ)
Video language: English

## Check the supported Mac setup

Thanks to Scrimba for sponsoring this video.

The July walkthrough uses apfel to expose Apple’s on-device foundation model. Its “latest macOS” wording is relative to the recording; check the actual OS, hardware, language and model requirements before installing.

1. Use the installation sequence shown at 1:31; the recorded command is brew install apfel.
2. Follow the first-run language/download check shown at 3:11.
3. Verify a basic reply before introducing private material.
4. Keep the recording’s small-context limitation in mind when choosing an example.

## Try the three modes separately

The commands shown in the source are apfel "hi", apfel --chat and apfel --serve. They serve different testing purposes.

1. Run the one-shot prompt and inspect the reply.
2. Start chat mode and check what context is available within that session.
3. Do not assume a new session remembers previous conversations.
4. Start the server route only for the intended local test and inspect its actual endpoint and access behaviour.

## Test a small script and its data path

The video adapts a Python client by changing its base URL and discusses an OpenAI-compatible local endpoint. That compatibility label alone does not prove every client feature works.

1. Use the actual local endpoint reported by the demonstrated server, not a guessed port.
2. Test a small input and inspect both errors and returned text.
3. Review the whole script for other network requests before calling its data flow private.
4. Choose short, checkable jobs and inspect the output against the input.

## The takeaway

This is a historical guide to the demonstrated local-model workflow. It does not establish current compatibility on every Mac or that every connected app keeps all data on-device.

## Workflow resources

- [scrimba.com](https://scrimba.com/?via=ai-luke)

## Chapters

- [0:00 Your Mac has a free, offline AI](https://www.youtube.com/watch?v=x87FIWoXIfQ&t=0s)
- [0:15 Apple quietly shipped on-device AI](https://www.youtube.com/watch?v=x87FIWoXIfQ&t=15s)
- [0:30 What we're setting up (2 minutes, works offline)](https://www.youtube.com/watch?v=x87FIWoXIfQ&t=30s)
- [0:50 How small is this model? (vs Haiku)](https://www.youtube.com/watch?v=x87FIWoXIfQ&t=50s)
- [1:13 The 4K context window catch](https://www.youtube.com/watch?v=x87FIWoXIfQ&t=73s)
- [1:31 Install it: brew install apfel](https://www.youtube.com/watch?v=x87FIWoXIfQ&t=91s)
- [1:49 What apfel actually is](https://www.youtube.com/watch?v=x87FIWoXIfQ&t=109s)
- [2:14 Sponsor: Scrimba](https://www.youtube.com/watch?v=x87FIWoXIfQ&t=134s)
- [3:11 First run + the language fix](https://www.youtube.com/watch?v=x87FIWoXIfQ&t=191s)
- [3:41 apfel "hi" — your first reply](https://www.youtube.com/watch?v=x87FIWoXIfQ&t=221s)
- [3:52 Chat mode for private data](https://www.youtube.com/watch?v=x87FIWoXIfQ&t=232s)
- [4:26 Why it has no memory](https://www.youtube.com/watch?v=x87FIWoXIfQ&t=266s)
- [4:55 Serve mode: a local API endpoint](https://www.youtube.com/watch?v=x87FIWoXIfQ&t=295s)
- [5:32 It's OpenAI-compatible (just change base_url)](https://www.youtube.com/watch?v=x87FIWoXIfQ&t=332s)
- [6:10 MCP server support](https://www.youtube.com/watch?v=x87FIWoXIfQ&t=370s)
- [6:17 apfel vs Ollama — which for what](https://www.youtube.com/watch?v=x87FIWoXIfQ&t=377s)
- [7:04 Parameter counts explained (3B active vs 33B total)](https://www.youtube.com/watch?v=x87FIWoXIfQ&t=424s)
- [8:01 The two real use cases](https://www.youtube.com/watch?v=x87FIWoXIfQ&t=481s)
- [8:24 Wrap-up + free Skool community](https://www.youtube.com/watch?v=x87FIWoXIfQ&t=504s)

## Sources and limitations

Written companion prepared 31 October 2026 from this exact published description and its chapter sequence where available. Checklists are editorial adaptations, not quotations from a reviewed final-cut transcript. Reported outcomes belong to the original video; the product was not retested. Model availability, setup details, prices and offers can change. Use the retained video and resources for the demonstrated setup.

[Original video](https://www.youtube.com/watch?v=x87FIWoXIfQ)