Skip to content

Kosmos-2

A multimodal LLM that links language with images

Visit official site

What Kosmos-2 is

The model is described as a multimodal large language model linking natural language with image and video inputs. The framework uses the GRIT dataset of image-text pairs for training and evaluation.

What you can do with it

  • Locate phrases in an image
  • Understand and generate referring expressions
  • Caption images and answer visual questions

Application scenarios

  • Generate an image description
  • Answer a visual question
  • Produce a referring expression for an object

What to watch out for

  • Listed tasks center on visual language
  • Training data is the GRIT dataset
  • Image and video are the named input forms

Pros & cons

✓ What we like

  • Image and video inputs are named
  • GRIT image-text pairs are used for training

! What to watch out for

  • Tasks center on grounding and visual language
  • Training data is described as the GRIT dataset

FAQ

What dataset is used?

The material names the GRIT dataset of image-text pairs.

What tasks are listed?

Phrase grounding, referring expressions, image captioning and visual question answering.

Last reviewed: 2026-11-15

More AI image generation tools

View all →

How we review