Downcodes editor reports: MiniCPM-V2.6, this 8B parameter multi-modal end-side model from OpenBMB, has recently achieved impressive results on GitHub and Hugging Face, ranking among the Top 3, and the number of GitHub stars has exceeded 10,000 , and the number of downloads exceeded one million! It comprehensively surpasses GPT-4V in terms of single image, multi-image, and video understanding, and for the first time integrates high-end functions such as real-time video understanding and multi-image joint understanding. Its efficient performance and convenient deployment method make it an important measure of the limit of end-side model capabilities, which has attracted widespread attention in the global technology circle.
Since its release, the latest version 2.6 of the MiniCPM-V series has quickly risen to Top 3 on the trend lists of GitHub and HuggingFace, the world's leading open source communities, and its number of GitHub stars has exceeded 10,000. Since its debut on February 1, the MiniCPM series has been downloaded more than one million times, becoming an important measure of the limit of client-side model capabilities.
MiniCPM-V2.6 achieves comprehensive performance improvements in single image, multi-image, and video understanding with its 8B parameters, surpassing GPT-4V. This end-side multi-modal model integrates high-end functions such as real-time video understanding, multi-image joint understanding, and multi-image ICL for the first time. It only occupies 6GB of memory on the quantized back-end side, and the end-side inference speed is as high as 18 tokens/s, which is 33% faster than the previous generation model. It supports llama.cpp, ollama, vllm inference, and supports multiple languages.
This technological breakthrough has aroused enthusiastic response in the global technology circle, and many developers and community members have shown great interest in the release of MiniCPM-V2.6.
At present, the GitHub and Hugging Face open source addresses of MiniCPM-V2.6 have been announced to the public, and links to llama.cpp, ollama, and vllm deployment tutorials have been provided.
MiniCPM-V2.6GitHub open source address:
https://github.com/OpenBMB/MiniCPM-V
MiniCPM-V2.6Hugging Face open source address:
https://huggingface.co/openbmb/MiniCPM-V-2_6
llama.cpp, ollama, vllm deployment tutorial address:
https://modelbest.feishu.cn/docx/Duptdntfro2Clfx2DzuczHxAnhc
The open source of MiniCPM-V2.6 provides developers with powerful tools and convenient deployment tutorials. It is believed that it will promote the development of end-side multi-modal model technology in the future and bring innovative possibilities to more application scenarios. The editor of Downcodes will continue to pay attention to its subsequent progress, so stay tuned!