From the source
Ant Ling released Ming-flash-omni-Preview, an open-source omnimodal large model with 103B total parameters and 9B activated, based on Ling 2.0's sparse MoE architecture.
It improves over Ming-lite-omni-1.5 in omnimodal understanding and generation, achieving a GenEval score of 0.90 and SOTA on all 12 ContextASR subtasks.
The model supports controllable image generation, streaming video understanding, speech and dialect recognition for 15 Chinese dialects, and voice cloning with continuous tokenizers.
