Ant Group introduces the Ling-3.0-Flash hybrid-attention model
First released July 23, 2026, Ling-3.0-Flash is a 124-billion-parameter mixture-of-experts model that activates 5.1 billion parameters per token and uses a 5:1 KDA-to-MLA ratio. Ant Group reports results comparable to or better than models two to three times larger. It offered free access through OpenRouter through August 3 and planned to publish weights afterward.
Visit original source