Local AI on Omarchy: setup, exact model and example prompts

What this setup covers

Run a downloaded language model in LM Studio on an existing Omarchy desktop. This is local text chat: rewriting, organising notes and planning an outline. It does not automatically give the model access to your files, desktop or coding tools.

This guide follows the documented route shown in the video. Luke has not installed or tested it on a running Omarchy machine. The fictional prompts below are inputs to try, not captured model results.

Install LM Studio

Open the Omarchy menu and follow Install → AI → LM Studio. Keep an internet connection during installation and initial model/runtime downloads.

LM Studio also supplies a Linux AppImage. Its documented Ubuntu guidance is not an Arch Linux compatibility guarantee; do not substitute an Ubuntu .deb into this route. x64 Linux requires supported hardware, with AVX2 part of the default documented requirements.

1. Check the current official Omarchy AI manual and LM Studio requirements before changing your system.
2. Use Omarchy’s AI menu on your existing installation, open LM Studio and let initial setup complete.
3. If you are using a virtual machine, check its assigned RAM and storage. GPU access depends on the VM setup; this video provides no GPU-passthrough or speed result.

Download one exact model file

Open Discover (Ctrl+2) and find lmstudio-community/Qwen3-4B-Instruct-2507-GGUF. Select Qwen3-4B-Instruct-2507-Q4_K_M.gguf, not every file in the repository.

The Q4_K_M download is approximately 2.5 GB. That is storage size, not a promise that the running model needs only 2.5 GB of RAM. Context length, runtime and GPU offload also affect memory use.

1. Download the exact GGUF file while online.
2. Switch to Chat, open the model loader and select the downloaded model.
3. Start with the default settings. Allow any required runtime to download while still online, then wait for the model to finish loading before sending a prompt.

Try a careful rewrite

Paste this literal English example into a fresh chat. You can request a different output language, but keep the facts and commitments unchanged.

Rewrite this fictional message in a friendly, clear tone, under 50 words. Keep every commitment accurate and add no promises: I'm still waiting for the files, so I can't finish the draft yet. I'll send an update tomorrow. Could you let me know when you expect to send them?

Check that the reply promises an update tomorrow, not a finished draft tomorrow. Reject extra deadlines, invented promises and unnecessary personal details.

Turn notes into a checklist

Use this fictional example; replace it with your own non-sensitive notes when you understand the workflow.

Turn these fictional project notes into a checklist. For each task, show its owner and deadline. Use "unspecified" when either is missing. Do not invent dates or commitments. Maya: revise the homepage copy by Thursday at 3 pm. Ben: provide three product screenshots by Wednesday at noon. The first draft is due Friday at 10 am. No owner has been assigned. The colour palette needs approval before the design can be finalised. No approver or approval deadline has been set.

Check Maya: Thursday, 3 pm; Ben: Wednesday, noon; first draft: Friday, 10 am, owner unspecified; palette: approver and deadline unspecified. Approval blocks final design. The model should not fill in those missing facts.

Plan a beginner video

This is an outline task, not evidence that a backup or product test has been completed.

Turn these rough notes into a five-part outline for a beginner video. Give each part a short heading and one sentence explaining what to show. Don't add product names, prices or claims about completed tests. Idea: backing up phone photos. Explain what the viewer wants to keep. Copy the photos to an external drive. Keep a second copy separately. Show why checking that a few files can be restored matters. Finish with a simple repeatable backup habit.

Inspect the order, the five-part limit and whether the outline preserves both a separate second copy and a restoration check. Remove invented product recommendations or claims of successful tests.

Check a fresh offline response

Finish downloading and loading the model first. Disconnect every internet route: Wi-Fi, Ethernet and the virtual network adapter if you are in a VM. Open a new chat (Ctrl+N) and send a new prompt.

Make a shopping list from these notes, grouped into fruit, vegetables and other items: apples, carrots, bread, bananas, onions, milk.

A fresh answer with all internet connections disabled supports this local-chat check. It does not audit the network activity of every app feature. Do not use a cloud model, web-search tool or connected MCP service for this check.

Troubleshoot and stop cleanly

If loading fails, close memory-heavy applications, try a fresh chat, reduce context length or choose a smaller compatible model. Record the operating system, app version, CPU/GPU, exact model file, settings and error before looking for a fix; do not install arbitrary drivers.

Unloading a model frees running memory but does not delete the downloaded file. Delete only the exact GGUF if you want to remove its stored copy. Chats may be saved as ordinary local files: consider who can access the device and whether backups copy them elsewhere.

Cloud subscriptions are optional for this local-chat path, but hardware, electricity, storage and other services can still have costs. Coding agents, desktop control and file access require separate configuration.

The takeaway

Start with one loaded model and one small text task. Check the facts in every answer before trusting it.