Yesterday, the world's strongest end-side multimodal model was refreshed again, with only 8B parameters, beating OpenAI's GPT-4V and Google's Gemini Pro, and its OCR long difficult image recognition refreshed SOTA, image encoding speed skyrocketed 150 times, and the world's strongest end-side multimodal model is not produced by foreign manufacturers, but the most head company in domestic large model research and development strength, Face Wall Intelligence, the latest wall small steel cannon MiniCPM-Llama3-V 2.5.
It is understood that the wall-facing small steel cannon MiniCPM-Llama3-V 2.5 only relies on the 8B end-side model, with a score of 65.1 on the evaluation platform OpenCompass, and the closed-source model Qwen-VL-Max may be able to fight, and the comprehensive performance directly beats the heavyweights GPT-4V and Gemini Pro. In the OCR comprehensive benchmark test, it achieved a score of 725 points, which is far beyond GPT-4V, and has also achieved significant improvements in the "stubborn" hallucination ability of the large model, as well as various benchmarks, the data is far better than GPT-4V and Gemini Pro.
To put it simply, the MiniCPM-Llama3-V 2.5 can see, read, be fast, think better, and leverage the strongest performance with the smallest parameters! So, putting aside these benchmarks, what can this little steel cannon bring to us ordinary people? First of all, it can support 30+ languages, including German, French, Spanish, Italian, Russian and other mainstream languages; Secondly, it supports the accurate recognition of difficult pictures, long pictures, and long texts, for example, if you are reading a long article that eats melons, you are always annoyed by "too long to read", and if you throw it to it, it can quickly summarize the key content; If it is an English version of graphic information, it can also give a very accurate summary according to your needs; Moreover, it can understand the subject content of the picture "at a glance" in a picture with a variety of element information, deduce the source information of the picture, and then organize and summarize the information after "thinking" to us. When it is installed on a mobile phone, it will work quickly and in real time directly on the device, reducing the risk of data breaches, and it can work even without an internet connection, and can be used on multiple devices, or become a true AI "personal assistant".