Lei Jun launches a fierce attack on AI large models

Author | Zhou Zhiyu

Xiaomi is launching a rapid offensive in the field of large models.

According to Wall Street News, Xiaomi, which has been very low-key in large models, has continuously increased its computing power reserves over the past few months and has also made plans for higher computing resource investments to provide more sufficient computing power supply for its large model research and development.

The further increase in capital expenditure on computing resources is a reflection of Xiaomi founder Lei Jun's aggressive approach towards AI large models. Previously, Xiaomi had already taken many actions in internal organizational capability building and external talent acquisition.

In mid-November this year, Xiaomi's Basic Technology Platform Department established the AI Platform Department, with Zhang Duo, who was publicly praised by Lei Jun as "Xiaomi's great god," serving as the head of the AI Platform Department.

Subsequently, one of the key developers of DeepSeek-V2, Luo Fuli, is also rumored to join Xiaomi, possibly joining the Xiaomi AI Laboratory. Luo Fuli is renowned in the field of natural language processing (NLP), especially for his involvement in DeepSeek-V2, which has attracted industry attention due to its significantly lower usage costs compared to industry averages. Luo Fuli's addition will also accelerate Xiaomi's research and development in the field of large models.

All signs indicate that under Lei Jun's leadership, Xiaomi is accelerating the progress of large model research and development. However, Xiaomi has been quite low-key in this area for some time.

In last year's annual speech, Lei Jun stated that Xiaomi would fully embrace AI large models. The Xiaomi AI Laboratory also established a dedicated large model team in April 2023.

People close to Xiaomi have indicated that the company is cautious about the need for large-scale spending on pre-training, while lightweight models have certain advantages over trillion-parameter large models in specific tasks. This has led Xiaomi to focus on "lightweight" and "local deployment" in its large model strategy.

Xiaomi's large model parameter scale is in the tens of billions, while for comparison, vivo launched its Blue Heart large model with a trillion-parameter scale in early November.

Xiaomi insiders believe that what sets Xiaomi apart from other companies is its emphasis on product implementation. This means that large models will be released alongside products.

Xiaomi Group President Lu Weibing has also stated that the currently released so-called AI phones are actually AI Feature phones, which means they have some AI functions created using AI technology, while a true AI phone would be one that runs an operating system reconstructed based on AI large models.

This line of thinking has led to a low level of awareness of Xiaomi's large models in the outside world.

At the end of this year, during the launch events of several smartphone manufacturers, the empowerment of their products' intelligence by large models became a key focus of the promotions. In contrast, during the launch event for Xiaomi's flagship phone, the Xiaomi 15, the emphasis was placed on the Xiaomi Surge OS 2.0, with no further detailed introduction to large models.

However, Xiaomi has made significant progress in developing its own large models. In May of this year, Xiaomi's large language model MiLM successfully passed the large model filing.

In November of this year, Xiaomi released the second generation model MiLM2 series, which has parameter scales ranging from 0.3B to 30B to meet the needs of various scenarios across cloud, edge, and terminal.

In terms of model scale, the MiLM2 series continues the lightweight approach, with parameter scales still in the tens of billions. The MiLM2-30B model is specifically designed for cloud scenarios and surpasses mainstream competing large models in instruction adherence, common sense reasoning, and reading comprehension In addition, as of mid-November, Xiaomi's total computing power for smart driving reached 8.1E FLOPS, placing it in the first tier among current automakers. The cumulative data accumulation has reached 3 million Clips, which is on par with Li Auto during the same period. According to Xiaomi's expectations, it will complete a data accumulation of 10 million Clips by the end of the year.

Of course, this still leaves a significant gap compared to Tesla's 100E FLOPS computing power. In the second half of the intelligentization of new energy vehicles, Xiaomi needs to continue to "maintain integrity while innovating" and accelerate its efforts in intelligentization. Therefore, it is not surprising that Xiaomi is further increasing its investment in computing power resources.

Compared to other tech giants, Xiaomi has a vast terminal ecosystem that includes smartphones, cars, and IoT. This will be an advantage as the AI large model sector experiences a competition among hundreds of models and enters a phase of seeking practical AI applications. However, this also requires Xiaomi to have more outstanding performance in the field of AI large models.

As Xiaomi intensifies its efforts in the AI large model sector, the competition for AI applications is also reaching a climax