Skip to content

Tarsier

A video language model for detailed video descriptions

Visit official site

What Tarsier is

The model is described as a series of large video language models from a ByteDance research team. It produces high quality video descriptions and works on video understanding tasks.

What you can do with it

  • Write detailed descriptions of video events
  • Examine the architecture and training method
  • Add automatic captions to a video service

Key features

  • Multi-task pretraining covers description and QA
  • Multi-granularity instruction tuning is applied
  • A DREAM-1K description benchmark is supplied

Getting started

  • Create a Python 3.9 virtual environment
  • Clone the repository and run the setup script
  • Download the Tarsier-7b or Tarsier-34b weights

Pros & cons

✓ What we like

  • The DREAM-1K description benchmark dataset is provided
  • Video, image and GIF inputs are supported

! What to watch out for

  • A Python 3.9 virtual environment and setup.sh are needed
  • Tarsier-7b or Tarsier-34b weights must be downloaded

FAQ

What training stages are used?

The card names multi-task pretraining and multi-granularity instruction tuning.

Which inputs are accepted?

The material lists video, image and GIF files.

Last reviewed: 2026-11-15

More AI video generation tools

View all →

How we review