China News Service, Beijing, November 11 (Reporter Liu Yuying) A number of mobile phone manufacturers in China have deployed AI large models in mobile phones or launched AI mobile phones. A number of experts interviewed said that the path of AI large models in mobile phones is clearer, but there are still some difficulties.
The AI model will become an important bargaining chip for mobile phone companies to compete in the future. Recently, Huawei, vivo, OPPO, Xiaomi and other manufacturers have launched mobile phone terminals equipped with large models, and all domestic head mobile phone manufacturers have laid out in this track.
At the OPPO developer conference held last week, its operating system ColorOS14 was officially unveiled. For the first time, the Andean model (AndesGPT) is built into this system update. It is understood that the model contains a variety of models with different parameter scales from 10 billion to <> billion, and the upgraded "Xiaobu Assistant" will have AI large model capabilities.
Last month, vivo launched a self-developed AI large model matrix, which includes five self-developed large models with three different parameter levels: billion, 5 billion, and 100 billion, covering core application scenarios. The company also released its first AI large model mobile phone X<> this month, which is the industry's first mobile phone with a <> billion large model in the terminal.
In addition, Lei Jun, chairman of Xiaomi Group, announced that Xiaomi has fully embraced the AI large model, and the large model on the mobile phone side has been preliminarily run. Huawei announced that the mobile phone system will be connected to the Pangu large model, and the public test of the voice assistant "Xiaoyi" with the built-in large model has been launched. Honor said that it will soon launch a self-developed 70 billion parameter device-side AI model and a new cloud service.
Zhang Yang, consulting director of Beijing Boruiheng Consulting Co., Ltd., said that at this stage, the application of mobile phone AI large models is mainly focused on image recognition, speech recognition, natural speech processing and other fields, which can provide consumers with more accurate language recognition, more intelligent information recommendation, smoother voice interaction, faster response speed and other experiences.
Liu Haifeng, President of OPPO's Digital Intelligence Engineering Division, said that with the support of generative AI technology, mobile phones will be more accurate in understanding user intentions and situational awareness.
The deployment of AI models in mobile phones is technically affected by technologies such as computing chips, heat dissipation, and batteries. Zhong Xinlong, director of the Artificial Intelligence Research Office of the Future Industry Research Center of CCID Think Tank, told China News Service that the recent upgrade of software and hardware technology has made the path of AI large models deeply landing on mobile phones gradually clear.
For example, one of the major selling points of the Qualcomm Snapdragon 10Gen26, which was just released on October 8, is that the offline terminal runs a large model with tens of billions of parameters. The Dimensity 3 processor released by MediaTek this month supports tens of billions of AI models, which can provide better games, images, and multimedia entertainment experiences. In addition, dedicated AI chips and hardware accelerators can significantly improve the efficiency and speed of model runs.
In reality, mobile phone manufacturers and software developers need to take an extremely complex approach to implement AI large models. Zhong Xinlong said that the combination of device and cloud is a common way to realize the application of large models on mobile phones. Computing tasks are distributed to cloud servers, and the phone handles some simple computing tasks locally, while sending more complex tasks to the cloud, which can be achieved by using smaller, optimized models or hardware acceleration.
In addition, the optimization of large models through model compression and optimization, such as knowledge distillation, model pruning, quantization, etc., can reduce the size of model parameters and computing requirements, so that it can run on mobile phones with limited computing power.
Unlike the landing of large AI models on cars or PCs, their landing on mobile phones has more challenges. Zhang Yang said in an interview that how to ensure the operation effect or ability of the large model on the premise of ensuring the battery life of the mobile phone, and how to ensure the security and privacy of user data are all challenges.
"At present, the application scenarios of mobile AI large models are still relatively limited, and whether there is enough market demand to support the application and promotion of this technology in the future remains to be further observed." Zhang Yang said.
The vice president of vivo said that there are three difficulties in mobile phone manufacturers to make large models in terms of data, algorithms and computing power, especially in the field of computing power, and the construction of large computing power clusters is challenging in itself. In recent years, vivo has invested more than 200 billion yuan in AI. (ENDS)