README.md

---
title: MOSS-VL-Realtime
emoji: 👁️
colorFrom: gray
colorTo: indigo
sdk: gradio
sdk_version: 6.15.1
app_file: app.py
short_description: Live streaming video chat — ask while the model watches
python_version: "3.12"
startup_duration_timeout: 1h
models:
  - OpenMOSS-Team/MOSS-VL-Realtime
---

# MOSS-VL-Realtime — Live Session

A realtime demo of [OpenMOSS-Team/MOSS-VL-Realtime](https://huggingface.co/OpenMOSS-Team/MOSS-VL-Realtime),
a multimodal vision-language model specialized for **realtime streaming video understanding**.

## How it works

1. **Stage a media** on the left — upload a video or image, or record from your camera
   (screen recordings can be uploaded as video files).
2. **Start a session** — the model watches the stream frame by frame at real playback pace.
3. **Chat while it watches** — type in the input box and press Enter; your question is pushed
   *into the running stream* and the model answers mid-stream. After the stream ends the
   session stays open for follow-up questions (until idle timeout or Stop).

The chat log shows one bubble per model response round, stamped with the video timestamp
at which it was produced. The `{ } Raw` toggle switches the log to the underlying event
stream (frames, silences, rounds, prompts).

One media per session — changing the staged media ends the live session, and starting a
new session clears the log (the finished transcript stays on screen until then).

## MCP

This Space is an MCP server exposing one tool, `analyze(media, prompt, ...)` — a stateless
single-shot analysis of a video/image path or URL.

## Local development

```bash
pip install "gradio[mcp]==6.15.1" pillow
MOSS_DEMO_MOCK=1 python app.py   # scripted model, no GPU/torch needed
```