Michael Limberger

In the room

This talk builds a character chatbot on your own hardware. We use Magidonia-24B through Ollama and OpenWebUI, and we keep the whole stack local.

The model for this write-up is Magidonia-24B-v4.3 by TheDrummer. In the 12B to 24B range it was the roleplay pick most people in that community reached for when these notes were written. We skip the basic install. You already have Ollama (the local model server) and OpenWebUI from earlier mixers. Plan on under two hours for character work, working examples, and parameter tuning.

The walk tonight

This tutorial takes you from a model already downloaded to a chatbot that stays in character and actually works. Here is the map:

• The model itself: why Magidonia was the pick in its weight class
• Configuration: the parameters that matter, and the ones that do not
• Character creation: how to write a system prompt Magidonia respects
• Working examples: prompts you can copy and take home
• Community patterns: techniques from the BeaverAI Discord (about 9,000 members when this talk ran)

You will see why each step exists, not only which button to press. The tone is coffee-chat: clear sentences, real explanations, no flyer slogans.

On the table

Hardware:

• Ollama installed (you have this)
• OpenWebUI running (you have this)
• 16GB+ RAM minimum (32GB+ recommended for 24B models without compromise)
• About 25GB free disk space

Time:

• About 90 minutes if you follow along
• About 30 minutes if you skip to the examples

Software:

- Same stack as January: Ollama + OpenWebUI

• No new dependencies
• Everything runs local

RPMax to Magidonia

If you attended January's RPMax tutorial, some pieces will look familiar, but the architecture choices are different.

RPMax uses ArliAI's base and chat template. Magidonia uses Mistral's architecture with TheDrummer's fine-tuning. RPMax was strong at instruction-following. Magidonia is tuned for creative work: roleplay, storytelling, and character consistency.

The communities differ too. RPMax lived in ArliAI spaces. Magidonia lived in BeaverAI, a large Discord where people merge, iterate, and talk about the model in real time.

One practical difference: for basic chat we do not need a custom Modelfile. The model ships with the Mistral V7-Tekken chat template baked in, and Ollama picks that up automatically. We used to write YAML for this. Now we usually do not.

Everything local

Everything stays on your machines. There is no cloud chat vendor in the middle, no subscription for the chat itself, and no third party listening in. You own the data and the chatbot. If something breaks, you fix it. If you want to change the model, you can.

That is the trade. You do more setup than ChatGPT, and you are responsible for the hardware. For many of us that trade is worth it.

Before we start

Grab a coffee and open a terminal. By the end of this series you should have a chatbot that can hold a conversation, stay in character, and remember who it is supposed to be.

Let's build something that actually runs on your desk.