// Popular Articles
Sapiens2: Meta vừa thả ViT backbone human-vision chất lượng cao nhất trong public domain
Meta công bố Sapiens2 tại ICLR 2026 — model ViT 0.1B–5B params, resolution 1K/4K, pretrained trên 750M ảnh người. Cộng đồng gọi đây là 'first public release by a large lab that is non-trivial to replicate'. Pose +4 mAP, body-part seg +24.3 mIoU, normal error giảm 45.6%.
Meta's REFRAG: 30× Faster RAG Decoding Without Losing Accuracy
Meta Superintelligence Labs just shipped REFRAG — a decoding framework that compresses RAG context into chunk embeddings, hitting 30.85× faster time-to-first-token, 16× longer context, and zero perplexity loss. No LLM retraining required.