AETHORIA AI

Open, efficient AI research

Community-led research into compact language models, token-identity hash routing, multimodal generation and accessible training systems.

Featured project

TR-HASH MoE 200M — Full SFT

Released full-parameter assistant checkpoint trained for three epochs on the audited 300K-example SFT v2 mixture, from the approximately 162B source-token lineage.

201.2M parameters 300K SFT examples Full-parameter SFT Released

Vision research

TR-HASH Vision v8

MoE-based hierarchical vision transformer, initialized randomly and trained end-to-end on COCO 2017 at 640 px, without external detector, classification backbone or pretrained weights. 50 epochs of pretraining, then 12 epochs of supervised refinement.

2.59M parameters COCO from scratch mAP50 32.5% mAP50-95 20.05%

Datasets →

Atlas Posttrain Text instruction and chat SFT corpus Atlas Images 336K provenance-aware image-text pairs Atlas Image Edits 336K instruction-guided editing triplets