Michael Limberger

Installation

This part is short. One pull command, a coffee while it downloads, then a quick check that the model is there.

The Pull Command

Open your terminal and run:

Show me

ollama pull hf.co/bartowski/TheDrummer_Magidonia-24B-v4.3-GGUF:Q8_0

This downloads straight from HuggingFace. Ollama can fetch GGUF files from HF when you prefix the model name with hf.co/. The download is about 25GB, so budget roughly 10 to 40 minutes depending on your connection. Grab a coffee or skim the BeaverAI Discord while it runs.

The model name is long on purpose. It encodes who quantized it, which model you got, and which quant level you picked:

hf.co/                              - HuggingFace, not Ollama's library
        bartowski/                          - The quantizer (bartowski uses iMatrix)
        TheDrummer_Magidonia-24B-v4.3-GGUF  - The model and format
        :Q8_0                               - The quantization level (8-bit, 0-variant)

When the download finishes, Ollama prints a digest (a hash) so it can verify the file was not corrupted. You do not need to do anything with that hash yourself.

Verify The Download

Once the pull completes, confirm the model is listed:

Show me

ollama list

You should see the full name and about 25GB of disk use. For more detail:

Show me

ollama show hf.co/bartowski/TheDrummer_Magidonia-24B-v4.3-GGUF:Q8_0

That shows parameters, metadata, and the embedded chat template.

Quick Test

Before opening OpenWebUI, smoke-test from the command line:

Show me

ollama run hf.co/bartowski/TheDrummer_Magidonia-24B-v4.3-GGUF:Q8_0 "Hello, who are you?"

The model should load, think briefly, and answer in plain, coherent language. Type quit and press Enter when you are done.

The Name Is Long

Typing that full name every time gets old. Later we create a shorter alias with a Modelfile. For now Ollama stores the full name internally, and the long string is just accuracy.

Alternative Pulls For Different Hardware

If your hardware differs, you can still follow along with another quant or sibling model.

For Q6_K (smaller, still strong):

Show me

ollama pull hf.co/bartowski/TheDrummer_Magidonia-24B-v4.3-GGUF:Q6_K

For the regular Cydonia variant instead:

Show me

ollama pull hf.co/bartowski/TheDrummer_Cydonia-24B-v4.3-GGUF:Q8_0

For a heretic build (fewer safety leftovers):

Show me

ollama pull hf.co/mradermacher/Cydonia-24B-v4.3-heretic-v2-i1-GGUF:Q8_0

Installation, verification, and the next steps are the same for all of these.

How Ollama Handles The Chat Template

Modern GGUF files often embed the chat template. Magidonia includes the Mistral V7-Tekken template, and Ollama picks it up automatically when the model loads. You do not need a custom Modelfile just to format system, user, and assistant turns.

In the January RPMax tutorial we wrote YAML Modelfiles for the template. Here Ollama does that for you, which removes a step. Magidonia was trained with this template, so it already understands it.

Open Web Ui: Immediate Availability

Once the model is pulled and OpenWebUI is running:

1. Open your browser to http://localhost:8501 (or wherever OpenWebUI is running)
        2. Look at the model selector dropdown at the top
        3. You'll see your new model in the list

You usually do not need to restart OpenWebUI. If ollama list shows the model, OpenWebUI should see it too because both talk to the same Ollama backend.

If it is missing, check:

Ollama is still running (ollama list works)
OpenWebUI is still reachable in the browser
A normal refresh, or a hard refresh (Cmd+Shift+R on Mac) if the model list is cached

Quick Test In Open Web Ui

Select the model from the dropdown and send something simple:

"Hello Magidonia, introduce yourself."

You may see a short pause while the model loads into memory, then a natural reply with no weird formatting. If that works, you are ready for parameters.

If it does not:

Check Ollama logs for memory errors or corruption
Confirm you still have 25+ GB free on disk
Retry the command-line test to isolate OpenWebUI from Ollama

Done

That is installation: one pull, one verification, one quick test. Next we tune parameters so Magidonia behaves the way you want for character chat.