README.md

---
title: AETHORIA AI
emoji: 🌌
colorFrom: indigo
colorTo: purple
sdk: static
pinned: false
---

<div align="center">

# AETHORIA AI

### Open research for efficient and accessible artificial intelligence

[Website](https://www.complexity-ai.fr/) · [GitHub](https://github.com/Complexity-ML) · [Discord](https://discord.gg/EyDqXqpxWu)

</div>

AETHORIA AI is a community-led, non-profit research initiative exploring how smaller and more efficient models can learn, reason and generate across text and images.

**Partner of [bench-labs](https://huggingface.co/bench-labs).**

## Research focus

- Small language models and efficient pretraining
- Token-identity hash routing and mixture-of-experts architectures
- Multimodal and generative models
- Original, traceable datasets and post-training methods
- Optimized training and inference on accessible hardware

## Our approach

We build experiments from first principles, publish reproducible artifacts and study architectures that can deliver useful capabilities without depending exclusively on massive models or inaccessible infrastructure.

Our work combines model design, dataset engineering, evaluation and inference systems. Projects are released progressively as their code, data and experimental evidence become ready for public use.

## Featured release

### [TR-HASH 0.5B Research Release](https://huggingface.co/collections/AETHORIA-AI/tr-hash-05b-research-release-6a78c87b936ff367bd77a9a1)

A token-identity hash-routed mixture-of-experts language model, with its base checkpoint, instruction-tuned release, post-training corpus and interactive demonstration.

### [TR-HASH Vision v6](https://huggingface.co/AETHORIA-AI/TR-HASH-Vision-v6-1M-COCO)

A compact hierarchical hash-routed detector with shifted-window attention,
native multi-stage features, residual level adapters, lightweight PAN fusion,
P2/STAL small-object supervision, and optional one-to-one NMS-free inference.
The complete detector will be initialized randomly and trained end-to-end
directly on COCO 2017 at 640 px with four hash-routed experts (top-2), MuSGD,
EMA, DDP and fused Triton kernels. No external detector, classification
backbone or pretrained weights are used. Final weights and reproducible COCO
metrics will be published only after validation.

[COCO 2017 from scratch](https://huggingface.co/AETHORIA-AI/TR-HASH-Vision-v6-1M-COCO) · [~1.31M parameters](https://huggingface.co/AETHORIA-AI/TR-HASH-Vision-v6-1M-COCO) · [4 experts · top-2](https://github.com/Complexity-ML/complexity-framework) · [MuSGD + EMA](https://github.com/Complexity-ML/complexity-framework) · [DDP + Triton](https://github.com/Complexity-ML/complexity-framework) · [Evaluation pending](https://huggingface.co/AETHORIA-AI/TR-HASH-Vision-v6-1M-COCO)

## Join the community

Follow our work on [GitHub](https://github.com/Complexity-ML), explore our releases on Hugging Face, or join the discussion on [Discord](https://discord.gg/EyDqXqpxWu).