Since its release, the latest version of the MiniCPM-V series, MiniCPM-V2.6, has quickly risen to the top 3 on both GitHub and HuggingFace trending charts, renowned global open-source communities. Its GitHub star count has surpassed 10,000. From its debut on February 1st to the present, the cumulative downloads of the MiniCPM series have exceeded one million, becoming an important benchmark for the limits of model capabilities on the edge.
MiniCPM-V2.6, with its 8B parameters, has achieved comprehensive performance improvements in single image, multi-image, and video understanding, surpassing GPT-4V. This edge-side multi-modal model is the first to integrate advanced features such as real-time video understanding, multi-image joint understanding, and multi-image ICL. It occupies only 6GB of memory after quantization on the edge, with an edge inference speed of up to 18 tokens/s, 33% faster than its predecessor, and supports llama.cpp, ollama, vllm inference, along with multiple languages.
This technological breakthrough has sparked a warm response in the global tech community, with many developers and community members showing great interest in the release of MiniCPM-V2.6.
Currently, the GitHub and Hugging Face open-source addresses for MiniCPM-V2.6 have been made public, along with links to deployment tutorials for llama.cpp, ollama, and vllm.
MiniCPM-V2.6 GitHub Open-Source Address:
https://github.com/OpenBMB/MiniCPM-V
MiniCPM-V2.6 Hugging 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