From the source
Ant Ling released Ming-Lite-Omni V1.5, an open-weight upgrade to its omni-modal model.
Built on Ling-lite-1.5, it has 20.3 billion total parameters with 3 billion active in its MoE section, and improves performance in image-text, document, video, and speech understanding, as well as image generation and editing.
The model achieves a GenEval score of 0.87 and a win rate of 87.07% against the previous version on internal evaluations.
