CONNECT WITH US

In China's battle for AI, Huawei hands in its results first while Xiaomi's LLM evaluation is revealed

Staff reporter, Taipei; Jack Wu, DIGITIMES Asia 0

Xiaomi's AI LLM "MiLM-6B" is revealed along with its benchmark evaluation data. Credit: AFP

Xiaomi recently revealed its LLM for the first time. Data from evaluation platforms C-Eval and CMMLU is revealed as well. Chinese smartphone brands are joining the LLM race one after the other. Huawei's smart voice assistant "Celia" has already launched, and Xiaomei just had an impressive debut. The AI battle among Chinese mobile phone makers seems to be on the horizon.

After Huawei, Oppo, and Meizu gave up on self-developed smartphone chips, it appears that smartphone makers are shifting their focus on the LLM generative AI trend. They are engaging in a new battle to push generative AI to the consumer market and allow consumers to experience this new technology on their mobile end devices.

Chinese smartphone brands like Xiaomi have previously disclosed their plans and progress in the LLM sector. Xiaomi founder and CEO Lei Jun and president William Lu have publically discussed Xiaomi's LLM and related strategies several times.

Recently, evaluation platforms C-Eval and CMMLU released scoring data for Xiaomi's LLM "MiLM-6B." Xiaomi's description of MiLM-6B on GitHub states that "MiLM-6B is a large-scale, pre-trained language model developed by Xiaomi. It has 6.4 billion parameters."

C-Eval and CMMLU are more well-recognized names in the field of Chinese LLM benchmark testing. They primarily assess the comprehensive knowledge reserves and language understanding capabilities of LLMs in the Chinese language domain.

In C-Eval's evaluations, MiLM-6B had an average score of 60.2, ranking tenth in the overall list and first in the same parameter level. It even ranked ahead of Alibaba Cloud's Tongyi Qianwen. In CMMLU's evaluations, MiLM-6B had average scores of 57.17 in Five-shot testing and 60.37 in Zero-shot testing.

In April 2023, Lei Jun wrote a post stating that Xiaomi would embrace large model technology. Later, during a conference call, Lu announced that the company had already established an AI lab large model team, consisting of over 1,200 professionals in the AI sector.

Lu once stated that Xiaomi won't do general large models like OpenAI but would instead combine them with the company's business to generate results through technology. Industry analysis believes that Xiaomi's LLM is expected to be applied to various scenarios, such as the smart personal assistant Xiao Ai, the mobile OS MIUI, smart cockpits, IoT, and robots.

As for Honor, CEO George Zhao mentioned at the Shanghai MWC that Honor has been collaborating with network service providers for its LLM demands. As of now, Baidu's ERNIE Bot, Alibaba's Tongyi Qianwen, and iFLYTEK's Spark are all LLM models launched by major internet and AI companies. Thus, seeking external collaborations will likely be Honor's approach.

As for Oppo, its team behind the "Breeno" smart voice assistant continues to conduct R&D for generative AI technologies. In April 2023, Alibaba Cloud announced that it's joining forces with Oppo's AndesBrain to build Oppo's LLM infrastructure. Tongyi Qianwen will be used as the foundation to complete the learning, fine-tuning, and front-end engineering for the LLM, creating an AI service for Oppo's end devices.

Whether it's Huawei, Xiaomi, Honor, or Oppo, Chinese smartphone brands are all expanding their IoT plans. The smart voice assistant has become the key entry point to link together all the equipment. Brands must incorporate LLM through voice assistants and then rapidly spread it to their entire software/hardware ecosystems.

Huawei appears to have taken an early lead, as its Celia smart voice assistant is already connected to its Pangu Model. It can write article summaries, send meeting invitations through email, and create personalized designs using photos.