Skip to content

Vary-toy

A compact Vary model built on Qwen-1.8B

Visit official site
Category AI models

What Vary-toy is

The model is described as a small Vary variant built on Qwen-1.8B with an improved visual vocabulary for wider generalization.

What you can do with it

  • Run document visual question answering on a GPU
  • Run chart question answering experiments
  • Run referring expression experiments

Model notes

  • Detection-driven positive samples replace natural-image negatives
  • 88.1% accuracy is reported on RefCOCO
  • 29% accuracy is reported on MMVet

Pros & cons

✓ What we like

  • Visual information about natural objects is encoded efficiently
  • Results are reported on four benchmarks

! What to watch out for

  • The base language model is Qwen-1.8B
  • Training is described for ordinary GPUs

FAQ

What is the base model?

The material names Qwen-1.8B as the base language model.

What ANLS is reported?

The material states 65.6% ANLS on DocVQA.

Last reviewed: 2026-11-15

More LLM API platform tools

View all →

How we review