Flow - 4x faster than typing
Mini toolsMOSS VL Realtime
By OpenMOSS
MOSS-VL-Realtime is the realtime streaming checkpoint of the MOSS-VL series, built for continuous video understanding rather than single pass offline analysis. It processes incoming frames continuously, lets users ask questions at any timestamp in a running stream, and answers using only the frames observed so far. The model can emit a silence signal when visual evidence is insufficient and revise earlier responses as new frames arrive. Each frame carries an absolute timestamp so the model can reason about event order, duration, and pacing. It uses a cross attention vision language architecture with Cross attention Rotary Position Embedding (XRoPE), mapping text and visual patches into a unified time, height, width coordinate space. It has 11B parameters, a 256K text context window, and supports image, video, and realtime streaming frame input with text output, including tool use through function calling.
New
Multimodal
Overview
MOSS-VL-Realtime is an 11B parameter vision language model built for continuous video streams. It perceives incoming frames and generates text in parallel, answers questions asked at any point in the stream, and can stay silent or revise earlier answers as new visual evidence arrives. It uses a cross attention architecture with timestamp aware frame encoding and a 256K context window.
Last updated: July 14, 2026
