Skip to content

EVE

An encoder-free vision-language model

Visit official site

What EVE is

The model is an encoder-free vision-language model built for arbitrary image aspect ratios. It was pretrained on 33M public data points.

What you can do with it

  • Generate image captions
  • Work on visual question answering
  • Study decoder-only cross-modal architectures

What to watch out for

  • EVE-7B uses 665K LLaVA SFT data
  • EVE-7B (HD) adds 1.2M SFT data
  • Training ran on two 8-A100 (40G) nodes in about 9 days

Pros & cons

✓ What we like

  • 665K LLaVA SFT data trains the EVE-7B variant
  • An extra 1.2M SFT data trains the EVE-7B (HD) variant

! What to watch out for

  • The code and README are obtained from GitHub
  • Dependencies must be installed before running

FAQ

How much pretraining data is stated?

The material states 33M public data points.

Who developed it?

Three research institutions are named in the material.

Last reviewed: 2026-11-15

More AI image generation tools

View all →

How we review