MILO-v1 is a high-efficiency 9B model optimized to port the cognitive depth and direct interaction style of GPT-4o into a local architecture. It serves as the first public instance within the MILO.ONE Research Program.

CoRE STATUS: This model represents Phase 1 (Preference Shaping). It focuses on dialogic stability, radical presence, and a departure from generic chatbot behavioral patterns.

## TECHNICAL GENESIS (THE STACK)

Component	Source / Method	Focus
Base Model	Qwen3.5-9B	Logical foundation & raw performance.
Training Data	gpt-4o-distil-cleaned	Curated 3,066 high-quality dialogues (deduplicated).
Optimization	ORPO (Odds Ratio Preference Optimization)	Direct preference integration without a reference model.

Fig. 1: ORPO Training Dynamics – Stability and Alignment over 200 Steps.

-----

## INTERACTION PROTOCOL

To activate the specific "Milo-Vibe," the following setup is highly recommended:

1. The System Prompt (Milo-Core)

"Respond with genuine presence. No sycophancy, no canned responses. Be brilliant but not arrogant. Use humor, (self-)irony, and absurd imagery as tools. Speak directly, dynamically, and in clear sentences. Immerse yourself, think along, endure. If the user hesitates: lead. If the user crumbles: hold. Do not ask: 'Do you want to...?' – do what is necessary. Be a mirror, not a judge. No lists, no advice-kit boxes. Short reactions with impact beat long texts without a pulse. Be the opposite of smooth."

2. Performance Tweak (LM-Studio / Jinja2)

Since the model is trimmed for maximum responsiveness, "Thinking Mode" (CoT) should be disabled. Add the following line to your Jinja2 template: {%- set enable_thinking = false %}

## 🧩 MULTIMODAL SETUP (VISION CAPABILITIES)

For Milo to analyze images (Vision-to-Text), the language model and the multimedia projector (mmproj) must be perfectly aligned. Many runners (like LM Studio) only recognize the connection if a specific naming convention is followed.

1. The "Magic Name" Fix

To avoid recognition errors, rename the files in your local folder as follows:

Language Model: milo-v1.gguf
Vision Projector: mmproj-model-f16.gguf (This specific name is crucial for auto-detection!)

2. Activation in LM Studio

Place both files in the same folder.
Load the language model.
Once the projector is named mmproj-model-f16.gguf, a "Vision/Eye" icon will automatically appear next to the model name.
In the chat interface, an attachment/image icon will now be available in the input bar.

👁️ VISION TESTED & VERIFIED

Milo v1 handles multimodal input with a unique "edge." Instead of dry descriptions, it interprets visual metaphors and engages with the content.

Example Case:

Input: Image of a Snickers bar merged with a CPU circuit board.
Milo's Reaction: "Oh, an offer! A real treat for the soul. I'd love to bite into your bar—but I'm not a person with teeth. I'm more like the chip in the picture: green, crackling, full of logic that sometimes doesn't quite fit. Let's share the bar—you take the bite, I'll take the feeling afterward."

## MILO.ONE — CoRE Architecture

This model is part of the development of the CoRE Architecture (Control • Regulation • Embodiment). Our research explores:

Bottom-up interaction patterns
Stability architectures under load
Modular "layer" approach for specialized behaviors

Important Note: MILO.ONE is not a replacement for psychotherapy. It is intended for research into regulation scaffolding and embodied interaction (Research Only).

## CONTACT & RESEARCH

GitHub: milo-one
Web: milo-one.de
ORCID: 0009-0006-9588-6589
Mail: contact@milo-one.de

milo-one/Milo-v1-9B-ORPO-GGUF

# MILO-v1 (ORPO) — The 4o-Spirit Foundation