Research release for TR-HASH 0.5B: base and SFT models, post-training data, demos, and supporting artifacts.
AI & ML interests
Small language models, token routing, mixture-of-experts, multimodal generation, efficient training, open datasets, multilingual NLP and optimized inference.
Recent Activity
View all activity
Organization Card
AETHORIA AI
Open, efficient AI research
Community-led research into compact language models, token-identity hash routing, multimodal generation and accessible training systems.
Partner of bench-labs ↗Featured project
TR-HASH 0.5B
Base and instruction-tuned models, post-training data and interactive inference demo.
492M parameters
TR-Hash MoE
20B pretraining tokens
Open artifacts
Vision research
TR-HASH Vision v6
Compact hierarchical hash-routed detector with native multi-scale features, shifted-window attention, residual level adapters and a lightweight PAN/P2 neck. The complete model will be initialized randomly and trained end-to-end directly on COCO 2017 at 640 px, without external detector, classification backbone or pretrained weights.
≈1.31M parameters
COCO from scratch
4 experts · top-2
P2 + PAN + STAL
DDP + Triton
Evaluation pending