Möbius

Bring Your Own Model → Self-Evolving AI

About • Architecture • How It Works • Quick Start • Contributing

---

About

Current LLMs update every few months. Between releases, they're frozen — no learning from interactions or real-time information. But humans don't work this way: our neural connections change through every experience.

Möbius is a system that makes LLMs continuously learn and self-iterate — serving and learning form one endless loop, like a Möbius strip. The model evolves from your interactions, and even from real-time information feeds.

For the philosophy behind this project — why we believe models should learn like humans, and how we plan to get there — see Proposal & Design.

Architecture

How It Works

1. User chats with AI
2. System captures conversations
3. Judge consistency between turns
4. Synthesize training samples
5. Train model with new samples
6. Model evolves

Quick Start

Prerequisites

• Python 3.10+
• Ray cluster
• GPU for inference and training

Installation

git clone https://github.com/vie-serendipity/Mobius.git
cd Mobius
uv sync

Running

uv run python -m mobius.main --ray-address ray://xx.xx.xx.xx:10001 --model Qwen/Qwen3-4B-Instruct-2507

Contributing

This is an early-stage project exploring continuous learning for LLMs.

If you're interested in this direction, feel free to open an issue or reach out to me via email.