MOSS-VL-Realtime
A vision model that knows when to shut up
What it is
OpenMOSS's 11B streaming vision-language model watches live video and generates text in parallel — you can interrupt with a question at any moment mid-stream. Cross-attention with XRoPE position embeddings unifies spatial-temporal reasoning across 256K context, and when the frames don't support an answer it emits a silence token instead of guessing.
Why it's interesting
Most VLMs answer after the video ends; this one answers during it, and the trained-in ability to deliberately stay silent when evidence is thin is a small design decision with big implications for robots, monitoring, and assistants that watch the world in real time. Apache-2.0 on both the weights and the MOSS-VL codebase.
Use cases
- Live camera-feed assistants that answer mid-stream
- Robotics and monitoring pipelines needing streaming perception
- Research on proactive vs. silent multimodal behavior
Who it's for
Vision researchers, robotics teams, realtime-app builders
Setup
Moderate. 11B in BF16 fits a single 24–48GB GPU; streaming inference runs through the MOSS-VL project stack
Limitations & cautions
The comprehensive technical report is still 'upcoming' and benchmark tables are sparse at check time — capability claims currently rest on the demo and the card. Young project, small community so far.
Editorial takeaway
Every demo reel shows a model that always has an answer. The interesting one is the model trained to say nothing until it's sure.