Lyre / documentation
Get Lyre
Providers and configuration

Run an AI model on your own computer

Download a free model, check that it replies, then use it for coding in Lyre. No cloud account needed.

On this page

How it fits together

Three pieces work together. You install the first one, download the second, and Lyre connects the third.

1 · Runner
An app that runs AI models on your computer, such as Ollama or LM Studio.
2 · Model
The AI itself: a file you download through the runner, usually a few gigabytes.
3 · Coding agent
The assistant in Lyre that uses the model to read and edit your project.
Good to know

Set this up on the computer that runs Lyre. Your phone or tablet uses that computer’s model through Lyre; it never downloads a model itself.

What you’ll need
  • Lyre open on your computer.
  • An internet connection for the downloads.
  • Free disk space and enough memory for the model and your other apps. The size is shown before you download; the recorded example model was about 2.5 GB.

Choose a runner

Pick one. If you’re not sure, choose Ollama: it has the fewest steps and Lyre finds it automatically.

RecommendedOllamaA small app plus simple terminal commands. Lyre can detect it and help download models for it.Most visualLM StudioBrowse, download and chat with models in a regular app window. No terminal needed.For tinkerersllama.cppTerminal tools for running model files directly. The most control, the most setup.

Install the runner you chose from its official site, then continue. Installing the runner and downloading a model are two separate steps.

Step 1 · Download a model and get a reply

Before involving Lyre, make sure the model works on its own. Choose your runner:

With Ollama

  1. Open a terminal

    On Windows open PowerShell. On macOS or Linux open Terminal.

  2. Download a small model

    Paste this and press Enter. The download can take a few minutes; wait until it says success. 

    ollama pull qwen3:4b-instruct

  3. Ask it something

    ollama run qwen3:4b-instruct "Say hello in one sentence."
    Terminal showing ollama pull completing with success, then ollama run replying Hello!
    A successful download and first reply.
Newer Lyre versions

The Local Models window can check for Ollama on this computer, link to the official download and, after you review the model’s license and size, download a recommended model for you. Lyre never installs Ollama itself.

With LM Studio

  1. Find a model

    Open LM Studio and go to Discover. Pick a small model; LM Studio shows whether it fits your computer.

  2. Download and chat

    Download it, load it in Chat, and send a short message such as “Say hello.”

  3. Start the local server

    Open the Developer tab and start the server. Lyre connects to this server, so leave it running. LM Studio’s instructions →

With llama.cpp

  1. Download a build

    Extract the official release for your computer from the llama.cpp releases page.

  2. Start the server with a model

    Run llama-server with a GGUF model file and a port, for example 8085. Keep that terminal open.

  3. Check it replies

    Open the address it prints in your browser and send a message. The full walkthrough has copy-ready commands for macOS, Linux and Windows.

It worked if

the model answered your message. If it didn’t, fix that first: Lyre can’t use a model that doesn’t reply on its own.

Step 2 · Add the model in Lyre

  1. Open local models

    In Lyre, go to SettingsAI & agentsConnections and choose Local models. Older layouts show Providers directly in Settings.

    Earlier Lyre Providers settings with the Local models button below the provider list
    Earlier recorded layout: the current Connections tab is inside AI & agents.

  2. Let Lyre find your model

    Lyre scans this computer for running model servers. Choose Add beside the model you downloaded.

    Add Local Model window listing two Ollama models found on the connected computer, each with Add and More buttons
    Found models appear with their size and where they run.

    Not listed? Make sure the runner (or LM Studio’s server) is running, then choose Scan Again.

  3. Choose Use Model

    When it shows Ready to use, select Use Model. The model now appears in the model picker for new conversations.

    Add Local Model window showing Ready to use for Qwen 3 4b Instruct with Use Model and Done buttons
    The saved connection is ready to select. Confirm a real reply in your next conversation.
Using llama.cpp or a custom port?

Under Don’t see your model? choose Connect to a local server, enter the server address (for example http://127.0.0.1:8085), and select Test Connection. When it shows Connected, add the model as above.

Connect to a local server form with address http://127.0.0.1:8085, Test Connection, Connected llama.cpp, and one model found
A manual connection to a llama.cpp server.

If Lyre asks to set up OpenCode, accept: it’s the coding agent that uses local models.

Step 3 · Try it on your project

  1. Start a new conversation

    Open a test project and start a new agent. Check that your local model is selected in the composer.

  2. Ask a read-only question first

    Read README.md and summarize it in three bullet points. Don't change files or run commands.
    A Lyre conversation where the local Qwen model read README.md and listed three categories
    A local model answering from the project’s own files.
You’re done when

the answer mentions what’s really in your file. Now you can ask for small edits and review each change before asking for the next.

Choosing a model your computer can handle

Start small. Choose a model that fits your available memory, then try it on a small task. Model size alone does not tell you how well it will code. Leave memory for Lyre, your project and the model’s conversation context.

  • Let Lyre suggest one. On newer hosts, the Local Models window recommends a model this computer can actually run, based on its memory.
  • Look for tool support. A model that chats well may still struggle to edit files. Coding agents need models that support tool use.
  • Context matters. Context is how much of your conversation and project the model can consider at once. Ollama recommends at least 64,000 tokens for agent work; check the allocation with ollama ps while the model is loaded. Ollama’s memory guide →
  • Check the license before using a model for commercial work.

If something doesn’t work

Lyre says no local models were found
Start the runner (for LM Studio and llama.cpp, the server) and choose Scan Again. A downloaded model that isn’t being served can’t be found.
Test Connection fails
Check the exact address and port printed by your server. Keep it on this computer (localhost or 127.0.0.1); connect your phone through Lyre, not directly to the server.
Replies are very slow, or it runs out of memory
Try a smaller model, close other memory-heavy apps, and keep tasks small.
It chats but can’t edit files
The model may not support tool use, or its context is too small for agent work. Try a model built for coding or tools.
Privacy

For the localhost setup in this guide, inference runs on the computer serving the model. Cloud models, web tools and other connected services you choose may still send data elsewhere. Model files belong to the runner: uninstalling or updating Lyre doesn’t delete them.

Want a longer tutorial with every screenshot? Read How to install open AI models on your PC.