Author | Zhou Zhiyu

Xiaomi is launching a vigorous offensive on the big model.

Wall Street News learned that Xiaomi, which is very low-key in its big model, has continued to increase its computing power reserves over the past few months, and also has a higher computing power resource investment plan to provide a fuller supply of computing power for the development of its big models.

Further increasing capital expenditure in terms of computing power resources is a reflection of Xiaomi founder Lei Jun's onslaught against AI models. Previously, Xiaomi had already taken many steps to build internal organizational capacity and bring in external talents.

In mid-November of this year, the Xiaomi Basic Technology Platform Department established the AI Platform Department. Zhang Duo, who was once publicly praised by Lei Jun as the “God of Xiaomi,” is the head of the AI Platform Department.

Subsequently, Luo Fuli, one of the key developers of DeepSeek-V2, was also rumored to be joining Xiaomi or the Xiaomi AI Lab. Luo Fuli is famous in the field of natural language processing (NLP). In particular, DeepSEEK-v2, in which she participated, attracted industry attention because the cost of using large models was far lower than the industry average. Luo Fuli's addition will also help Xiaomi accelerate research and development in the field of large models.

Judging from various signs, under Lei Jun's leadership, Xiaomi is speeding up the development of large models. However, when it comes to big models, Xiaomi has kept a pretty low profile for some time.

In his annual speech last year, Lei Jun said that Xiaomi would fully embrace the big AI model. The Xiaomi AI Lab also set up a dedicated big model team in April 2023.

People close to Xiaomi said that Xiaomi is cautious about pre-training, which requires large-scale money-burning, and that lightweight models also have certain advantages over 100 billion models in certain tasks. This also made Xiaomi focus on “lightweighting” and “local deployment” on larger models.

The parameter scale of Xiaomi's large model is 10 billion. As a comparison, the big blue heart model launched by vivo in early November already has a model in the order of 100 billion.

People involved in Xiaomi believe that the difference between Xiaomi and other companies is that it focuses on product implementation. This allows the big models to come out with the product.

Lu Weibing, president of the Xiaomi Group, also said that the so-called AI phones currently released are all AI feature (feature) phones, which use AI technology to perform some AI functions, and the real AI phones are equipped with an operating system restructured based on the AI model.

This kind of thinking has made the big model of Xiaomi not very well recognized by the outside world.

At the press conferences of many mobile phone manufacturers at the end of this year, the intelligent power of big models to their own products became the focus of publicity at the press conference. Xiaomi, on the other hand, focused on the Xiaomi Surge OS 2.0 at the launch of its flagship phone, the Xiaomi Mi 15 this year, but there was no detailed introduction to the big model.

However, the big model developed by Xiaomi has progressed quite a bit. In May of this year, Xiaomi's big language model MiLM successfully passed the Big Model Filing.

In November of this year, Xiaomi's second-generation model MiLM2 series was released. It has multiple parameter scales of 0.3B to 30B to meet the needs of various scenarios on the edge of the cloud.

Judging from the model scale, the MiLM2 series also continues the lightweight idea, and the parameter scale is still 10 billion. Designed specifically for cloud scenarios, the MiLM2-30b model surpasses mainstream competitor models in terms of command following, common sense reasoning, and reading comprehension.

Furthermore, as of mid-November, the total computing power of Xiaomi Smart Driving reached 8.1E FLOPS, which is currently in the first tier among automakers. Furthermore, the cumulative data accumulation reached 3 MillionClips, which is ideally in the same tier as the same period. According to Xiaomi's expectations, it will complete 10 MillionClips data accumulation by the end of the year.

Of course, this is still quite far from Tesla's 100E FLOPS computing power. In the second half of intelligent new energy vehicles, if Xiaomi wants to continue to “keep the right and act surprisingly,” it also needs to accelerate intelligence. As a result, it is no surprise that Xiaomi has further increased its investment in computing power resources.

Compared to other major technology companies, Xiaomi has a vast terminal ecosystem, covering mobile phones, automobiles, and IoT. This will be an advantage when the big AI model is going through the war of models and entering the stage of finding the implementation of AI applications. But this also requires Xiaomi to have a more prominent performance in the field of large AI models.

As Xiaomi expands the field of AI models, this AI application dispute is also gradually reaching a climax.