Spaces:
Running
Running
| 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. | |
| ## 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 MoE 200M — Full SFT](https://huggingface.co/AETHORIA-AI/TR-HASH-MoE-200M-160B-SFT) | |
| A released 201.2M-parameter deterministic token-routed MoE assistant. The | |
| root checkpoint is a full-parameter SFT, trained for three epochs on the | |
| audited 300K-example SFT v2 mixture from the approximately 162B source-token | |
| lineage. | |
| ### [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). | |