MOSS-VL Technical Report

Summary

MOSS-VL is an open vision-language model family that treats real-time interaction—perceiving while speaking—as a first-class capability. It is co-designed across the stack: the language decoder attends to vision only through gated cross-attention, enabling natural perception of incoming frames during generation; a synthesized interaction corpus supervises when to speak, stay silent, or revise; and a staged curriculum concentrates real-time-specific training in a light final stage over a strong offline foundation. Offline, MOSS-VL-Instruct is competitive at comparable scale and leads temporal-reasoning video sets. Across four streaming benchmarks, MOSS-VL-Realtime achieves the best average on three (second on the fourth) among open-source streaming models, sweeping the three subsets that squarely test proactive behavior—e.g., 66.0 vs. 37.5 for the best baseline on OmniMMI Proactive Alerting. With 11.3B parameters but visual tokens outside the decoded sequence, MOSS-VL widens its time-to-first-token advantage over same-backbone Qwen3-VL-8B from 2.8x to 5.1x as visual context grows. All five checkpoints, the training curriculum, and real-time inference code are released.