LiveCC

Streaming video LLM trained on ASR transcripts at scale (CVPR 2025): a novel streaming training approach densely interleaves ASR words with video frames, teaching temporally aligned, real-time video commentary. Ships the Live-CC-5M pretraining and Live-WhisperX-526K SFT datasets plus 7B models — the 7B set SOTA for real-time commentary among open models at release and stayed competitive on standard video QA.

From Mike Zheng Shou's ShowLab. Models are built on Qwen2-VL backbones, not from-scratch pretrains.