Skip to content

InternVL2_5-2B-MPO

A multimodal model family built on preference optimisation

Visit official site

What InternVL2_5-2B-MPO is

The model is described as a multimodal large language model family built on InternVL2.5 and mixed preference optimisation, joining InternViT with pretrained language models.

What you can do with it

  • Feed multiple images or video data
  • Split images into 448x448 pixel tiles
  • Learn relative preferences between responses
  • Load the model with transformers
  • Run in bf16, fp16 or 8-bit modes

Model notes

  • An MLP projector is randomly initialised
  • InternLM 2.5 and Qwen 2.5 are named as language models
  • MMBench v1.1, MMStar and MMMU are listed as benchmarks

Getting started

  • Install torch and transformers
  • Load the model and tokenizer
  • Prepare image and text inputs
  • Adjust inputs for single image, multi-image or video

Pros & cons

✓ What we like

  • A quick start guide uses the transformers library
  • Mixed preference optimisation learns relative response quality

! What to watch out for

  • Inputs follow the supported image and video modes
  • Loading relies on torch and transformers

FAQ

Which benchmarks are named?

MMBench v1.1, MMStar and MMMU are listed.

How is it loaded?

The steps use AutoModel and AutoTokenizer from transformers.

Last reviewed: 2026-11-15

More LLM API platform tools

View all →

How we review