AI stands as the true heart of Xiaomi's future endeavors.

Recently, news of Lei Jun's multi-million-yuan salary offer to lure Luo Fuli, a key developer from DeepSeek, has sparked widespread discussion in the AI market at the end of the year. Notably, Luo Fuli, the central figure in this tale, is not even 30 years old and epitomizes the post-95s generation.

The combination of being a post-95s and commanding a multi-million-yuan salary is rare globally. According to reports from Sina Finance and other media outlets, Luo Fuli will directly become one of the leaders of Xiaomi's large model team. Xiaomi hopes she can leverage her experience in developing DeepSeek-V2 to optimize and enhance the efficiency and user experience of their AI large model.

Various signs indicate that Xiaomi is highly determined this time around and has made AI a new core project to promote. In 2024, companies like Huawei, OPPO, vivo, Honor, Samsung, and even Apple have all embraced AI as their core strategies. Some have launched AI phones, some have introduced AI OS, and others have rolled out AI agents. However, Xiaomi has been observing or holding back, neither confirming nor denying the concept of AI phones, and the company has not proposed an AI strategy at the corporate level. Now it seems that Xiaomi is finally poised to make a significant move in AI in 2025.

With Lei Jun personally leading the charge, has the wind of large models finally reached Xiaomi?

Regarding the DeepSeek AI large model that Luo Fuli previously oversaw, our Lei Technology has published an article providing a detailed explanation. Its computing power price is as low as 0.1 yuan per million tokens (cache hit), while the cache miss and output prices are as low as 1 yuan and 2 yuan per million tokens, respectively. When compared to the prices of other AI large models, it becomes evident that DeepSeek is the most cost-effective among its peers.

(Image source: Lei Technology production)

In other words, if DeepSeek isn't subsidizing users at a significant loss, it must have made groundbreaking progress in optimizing computing power costs. Even the training cost for the DeepSeek-V3 large model was only $5.576 million. From both application and training perspectives, DeepSeek can be considered a paradigm of the "economical" AI large model.

This also happens to address Xiaomi's weakness. As the mobile phone brand with the largest product portfolio (including cars, Mijia products, etc.), cost is Xiaomi's primary consideration if it aims to popularize AI across more hardware ecosystems. After all, most consumers are reluctant to pay an additional monthly fee for a single AI function, so this cost must be borne by the enterprise itself. Improving model efficiency becomes paramount.

However, some media outlets have criticized Luo Fuli's move to Xiaomi, stating that "the wind of AI large models has finally reached Xiaomi," which I believe is somewhat biased. After all, Xiaomi released its first self-developed large model as early as August 2023 and successfully implemented end-side deployment. Xiaomi's AI team was established as early as 2016, making "Xiaoai Classmate" one of the most experienced smart voice assistants.

To propel the development of AI large models, according to reports from Jiemian News and other media, Xiaomi is actively constructing its own GPU cluster with ten thousand cards, with Lei Jun personally overseeing the entire project team. The current scale of relevant personnel in this project has surpassed 3,000, indicating that the importance of AI projects within Xiaomi is second only to its mobile phone and automotive core segments. Theoretically, AI will also become the new cornerstone of Xiaomi's mobile phones and automobiles (and even all its businesses).

As the absolute core of the future smart ecosystem, AI deserves such meticulous attention from Xiaomi and Lei Jun. In Lei Technology's view, Xiaomi is currently lagging behind other brands.

Xiaomi AI: An Early Riser but a Late Arriver

Xiaomi can be considered one of the earliest brands in China to heavily invest in the AI field, but it has been unusually slow in implementing AI large models and AI applications. Super Xiaoai has been officially announced for over a year, yet it has yet to deliver any truly revolutionary functions. If you interact with Xiaoai, you'll notice that it has indeed become quite chatty.

As a fan who was among the first to purchase the Xiaomi 15, I noticed a dedicated submenu in the settings of Pengpai OS 2.0 that introduces the currently available AI functions in the system. Although it lists 10 different functions, a closer look reveals that many of them are quite common, such as AI writing, AI conversation translation, and AI real-time subtitles.

Image source: Lei Technology

Why are these functions so common? Because you can find them in any AI app you download, and there won't be much difference in the user experience. Among all AI functions, only a few, such as AI anti-fraud and active visual perception (quick QR code scanning), are based on the local system.

Taking "active visual perception" as an example, from its initial introduction, this function relies on local AI image recognition and can quickly scan QR codes in the desktop state. During my experience, the perception and transition speed were not slow, and it would jump to the browser or preset payment software page based on the QR code type, making it the only function that provided me with a genuine "AI experience".

However, I found that many Xiaomi users were unaware of this function and couldn't find it on the AI function page. That's correct because "active visual perception" is not located on the secondary page of Pengpai AI but is placed in "More Settings." So, what's on the Pengpai AI page? The answer is nothing; it's all function introductions, giving users a false impression.

Image source: Lei Technology

To put it bluntly, if the programmers had added buttons on the Pengpai AI page that directly linked to the corresponding AI functions, I would have considered their efforts commendable.

Certainly, the more intriguing part was Lu Weibing's annual summary livestream some time ago. During this livestream, the experience leader of Pengpai OS was invited on stage to demonstrate some new AI functions of Super Xiaoai, but almost all the new functions encountered issues.

For instance, the first demonstration involved "double-tapping the white bar at the bottom to wake up Super Xiaoai." The leader couldn't wake it up by double-tapping the white bar with different gestures and even nervously wiped his hands on his pants several times. Ultimately, after numerous attempts, the function couldn't be reproduced during the livestream, setting the stage for subsequent problems.

Image source: Lei Technology

Using voice commands to wake up Xiaoai and set navigation based on locations displayed on Dianping resulted in several misunderstandings before finally succeeding; using the voice command "I want to take a selfie" to start the camera mode and then asking Xiaoai to caption the photo and post it to Weibo either resulted in no response or launched WeChat instead.

It's hard to imagine how much credibility Pengpai OS would retain if users experienced such AI functions. I'm afraid the system leader would once again face overwhelming criticism from Xiaomi fans.

Xiaomi AI Closes the Stable Door After the Horse Has Bolted, But It's Not Too Late

At least based on the current experience, it's difficult to say that Xiaomi AI has anything particularly attractive, whether it's already pushed functions or functions under development. For example, automatically jumping to Gaode Navigation based on the address displayed on the page only optimizes the function of "navigating to xxxx (specific place name)" to "navigating to this address." It's hard to argue that such an upgrade warrants the use of AI for optimization.

Another example is using voice commands to start a selfie and post it to Weibo. I understand that this function is designed for users who love to share, but the developers seem a bit too idealistic. Assuming a user loves taking selfies and is enthusiastic about sharing photos on social networks, what would she do after taking a photo? I asked two friends, and both replied: first, edit the photo, then select the appropriate photo, and finally add a caption.

For "social butterflies," their intention in sharing is to let friends and netizens understand their status, and it would be even better if they could receive some praise. For this reason, they would spend ten to fifteen minutes editing and captioning photos. Having AI randomly take a photo and directly post it to social networks is a classic case of "not quite right" and fails to truly please the target audience.

Therefore, I also asked them what AI functions they would prefer. The answer was AI automatic photo editing, AI automatic collage selection (choosing the right photos), and AI automatic captioning. It would be perfect if content could be generated based on users' writing habits. As for the current function demonstrations, to be honest, they seem more tailored for "straight men".

For comparison, let's consider the progress of other mobile phone brands in AI functions. vivo has made considerable strides, such as the previously launched Four Seasons Portrait function, which can alter the theme of images through AI, like filling the environment with falling leaves or snowflakes, and this function is actually applicable to all photos.

Image source: Lei Technology

Additionally, vivo supports Super Semantic Search, allowing users to directly search for photos that meet their requirements through voice commands, such as "find a photo taken in Beijing." In contrast, Xiaoai's response was to tell me places suitable for taking photos in Beijing, which was somewhat disappointing for me as a Xiaomi fan.

Image source: Lei Technology

Besides vivo, Honor's AI assistant can even order takeout with a single sentence. During the launch event of the Honor Magic7 series, it was demonstrated live by ordering 2,000 cups of Luckin Coffee. Whether from a practicality or user experience perspective, other brands have surpassed Xiaomi in exploring the realm of AI. It's no wonder Lei Jun personally leads the AI project because if Xiaomi doesn't "go all in," it will only fall further behind.

However, Xiaomi actually possesses its own advantages. Its vast smart ecosystem provides a significant edge in user data accumulation and can disseminate AI throughout the entire IoT ecosystem, further integrating AI with people, cars, and homes—an advantage other brands lack.

Whether Xiaomi can fully leverage this advantage and create AI functions that better cater to user needs will be the focus of Xiaomi's AI team moving forward. The current AI mobile phone ecosystem is still in its nascent exploratory stage. Even brands like Honor and vivo, which are performing well, are only slightly ahead. For Xiaomi, there is still an opportunity to catch up. I also hope that Jin Fan and the Pengpai OS team won't disappoint Xiaomi fans again.

Source: Lei Technology