---
title: Marlin 2B Video Understanding
emoji: 🎬
colorFrom: blue
colorTo: indigo
sdk: gradio
sdk_version: 6.10.0
app_file: app.py
short_description: Dense video captions and timestamp search
python_version: "3.10"
startup_duration_timeout: 1h
models:
  - NemoStation/Marlin-2B
---

# Marlin 2B Video Understanding

ZeroGPU Gradio demo for [NemoStation/Marlin-2B](https://huggingface.co/NemoStation/Marlin-2B), a 2B video VLM for dense video captioning and natural-language temporal grounding.

The app exposes two model-card workflows:

- **Caption**: returns Marlin's parsed scene paragraph plus timestamped events.
- **Find**: resolves an event query into a parsed start/end time span.

The model is loaded once at module startup and inference runs through a single `@spaces.GPU` endpoint. Example videos are short, attributed clips intended to keep verification fast while still exercising video decoding, caption parsing, and span parsing.