MOSS-VL
A series of open-weight 11-billion-parameter vision-language models for long-form and real-time video understanding, including a variant that watches a continuous video stream and decides on its own when to respond.
MOSS-VL is a series of three open-weight, 11-billion-parameter models for video understanding, built on a shared cross-attention architecture that decouples visual encoding from language reasoning. MOSS-VL-Realtime handles continuous video streams, running perception and text generation in parallel so it can be interrupted at any moment, decide on its own when to answer or keep watching, and correct earlier answers as new frames arrive. MOSS-VL-Instruct is built for offline use with strong long-video understanding and dialogue, and MOSS-VL-Base is a pretrained foundation for continued pretraining and fine-tuning.
Unlike models that watch a full clip before answering, MOSS-VL-Realtime targets multi-turn, low-latency interaction with fine-grained timing of when events occur in the video and streamed responses.
It is written mainly in C++ and distributed under the Apache-2.0 license.