Skip to content

MiniCPM-Llama3-V 2.5

An 8B multimodal model built for on-device use

Visit official site
Category AI models

What MiniCPM-Llama3-V 2.5 is

MiniCPM-Llama3-V 2.5 is described as the latest on-device multimodal model from the OpenBMB project, with 8B parameters and support for more than 30 languages. The model uses quantisation, CPU, NPU and compiler optimisation for efficient deployment on terminal devices.

What you can do with it

  • Run image and text interactions on a smartphone
  • Recognise scene text and extract information
  • Hold cross-language conversations and generate content
  • Deploy on terminal devices through quantisation

Model notes

  • 8B parameters are stated in the introduction
  • OpenCompass average score is given as 65.1
  • OCRBench score of 725 comes with full-text OCR extraction
  • RLAIF-V alignment is stated to lower hallucination rates

Getting started

  • Clone the OpenBMB/MiniCPM-V repository to a local machine
  • Create a conda environment and install the dependencies
  • Run the local WebUI Demo for the device type, such as NVIDIA GPU or Mac MPS
  • Fine-tune with Huggingface Transformers or the SWIFT framework

Pros & cons

✓ What we like

  • Quantisation and compiler optimisation target terminal devices
  • RLAIF-V alignment is stated to lower hallucination rates

! What to watch out for

  • Deployment steps require a local conda environment and dependencies
  • Fine-tuning relies on Huggingface Transformers or the SWIFT framework

FAQ

How large is the model?

The material states 8B parameters and support for more than 30 languages.

What scores are given?

The feature list reports 65.1 on OpenCompass and 725 on OCRBench.

Last reviewed: 2026-11-15

More LLM API platform tools

View all →

How we review