Open MOSS Open-Sources Real-Time Multimodal Model

AdinaYakup · x · 2026-07-19

Open MOSS has open-sourced **MOSS-VL-Realtime**, which is now available on Hugging Face. This 11B model family supports text, single/multiple images, single/multiple videos, and interleaved image-text inputs, covering both Chinese and English scenarios. The project highlights several design choices: using **Cross-Attention** to separate visual encoding from language reasoning, applying **XRoPE** for unified spatiotemporal positional encoding, providing a unified conversation template applicable for offline/streaming/real-time interactions, and supporting a **256K token** context. It can also continue processing new frames while generating a response, modifying or interrupting its answer based on scene changes, and even remaining silent when evidence is insufficient.

Original post →

More from Multimodal

Multimodal channel →