Qwen Team Releases OmniVChat: Native Audio-Video Dialogue Task, Bench and RL Recipe in One Drop

Alibaba's Qwen team posted OmniVChat, a package covering native audio-visual dialogue where omni models take simultaneous voice-plus-video input and reply in text without intermediate ASR or captioning. The release bundles OmniVChat-Studio, a multi-agent data engine that synthesizes single- and multi-turn dialogues, OmniVChat-Bench across five capability axes with a human-recorded validation subset, and an RL reward design targeting correctness, efficiency and dialogue naturalness. Paper hit Hugging Face's daily list with 30 upvotes shortly after release.