MOSS-VL

An open-weight 11B vision-language family for long-form and real-time streaming video understanding. MOSS-VL-Realtime perceives an incoming video stream while generating text, supports interruptible multi-turn dialogue, and autonomously decides when to speak or continue watching. Offline Instruct and Base checkpoints share the 256K context and are aimed at long-video reasoning and continued training.

The architecture connects a Qwen-family language backbone to vision through cross-attention rather than interleaving visual tokens in the text sequence. Its Cross-attention RoPE (XRoPE) places text tokens and visual patches in a shared time–height–width coordinate system; absolute timestamps support second-level temporal grounding. The July generation reports open-weight SOTA across OmniMMI proactive alerting, StreamingBench proactive output, and ProactiveVideoQA. Apache 2.0 weights are available, including August 11 FP8 and NF4 builds sized for a single 24 GB GPU.