Photo: The reporter provided the picture of “Robot Dream”, and MiniCPM gave a detailed description. \Photo by Guo Hanlin, a reporter from Ta Kung Pao

[Ta Kung Pao News]The global artificial intelligence big model has entered the era of “hundred-model war”. How to make it enter the homes of ordinary people as soon as possible has become a major proposition. The big model MiniCPM-Llama3-V2.5 launched by Mianbi Intelligence has excellent character recognition capabilities (OCR), can accurately recognize difficult and long images, and can run on terminals such as smartphones without connecting to the Internet. On June 5, Mianbi Intelligence and Tsinghua NLP Laboratory, after serious discussions, decided to make Mianbi’s “small steel cannon” MiniCPM free for commercial use.

In order to better understand the functions of the model, a reporter from Ta Kung Pao downloaded the model from the ModelSocpe community and tried it out. Imitating the official practical cases, a high-speed rail ticket was inserted into the model. Even though the picture clarity was low, the model was still able to give accurate answers and present a specific format through instructions to inform all the text information of the ticket stub. In addition to text recognition, the MiniCPM-Llama3-V2.5 model is also very accurate in image processing. When the reporter put in a picture from the movie “Robot Dreams”, although the model could not provide specific character names and picture sources, it was able to vividly summarize the entire content of the picture as “anthropomorphic puppies and robots in cartoons.”

However, even though the model is powerful, some details still need to be improved. The reporter observed in the experience that the model would appear to be “generated out of thin air” (i.e. “AI hallucination”) when processing a large amount of information that needs to be analyzed and processed. When uploading a promotional poster for the TV series “Kuang Bi” and asking it to recognize all the names in the picture, the names of actors who did not participate in the film, such as Sun Honglei, appeared.

It is reported that MiniCPM has been running on mainstream international mobile phone brands and terminal CPU chips, and can run smoothly even on older models that have been released for many years. In the view of Dr. Liu Yi, founder of Beike Ruisheng and a specially appointed expert of the National Major Talent Program, terminals such as computers, mobile phones, and watches are the information portals closest to users. After being combined with large models, terminal devices can be more flexible and intelligent, becoming real “assistants”, which will accelerate the popularization of AI technology.