In a single holiday season, the global AI landscape has undergone a seismic shift due to the emergence of DeepSeek. The impact of this Chinese startup and its groundbreaking product continues to reverberate throughout the tech world.

On February 5, the popularity of the DeepSeek concept propelled nearly 20 Chinese A-share companies to their daily trading limits. Among them, Parallel Technologies, listed on the Beijing Stock Exchange, closed at a 30% daily limit, while Daily Interactive maintained a 20% daily limit. Conversely, the U.S. stock market has fared less favorably. Since January 20, NVIDIA's market capitalization has declined by 13.84%, TSMC by 3.52%, and Google and AMD, despite strong financial growth, have not been well-received by Wall Street.

DeepSeek's rise has not only bolstered the confidence of China's tech companies but has also sparked renewed optimism in Europe, which believes it "still holds the potential to become a key player in AI innovation."

Amidst the excitement, however, Kimi, which launched a new model concurrently with DeepSeek, appears to be overshadowed.

Is Kimi falling out of favor?

Last year, "everyone was talking about Kimi," but this year, the conversation has shifted to "everyone is talking about DeepSeek."

As the Spring Festival approached, domestic large models rushed to update before the holiday. Alongside DeepSeek's release of its DeepSeek-R1 model, Dark Side of the Moon introduced its new Kimi k1.5 model. Both inference models are fully comparable to OpenAI GPT-4. Initially, both DeepSeek and Dark Side of the Moon's new models garnered significant buzz among overseas users upon their launch. However, DeepSeek's rapid ascent in the AI sphere quickly shifted the global tech spotlight almost entirely onto it.

In overseas markets, mainstream European and American media outlets worked tirelessly over weekends and reported overnight. They described the impact on the U.S. industry as "shocking Silicon Valley," and NVIDIA, rarely issuing such statements, praised the R1 model as "an excellent advancement in AI."

As DeepSeek's popularity soars, its user base is growing at an accelerated pace.

According to the latest statistics from the domestic AI product ranking, DeepSeek achieved remarkable results just 20 days after its launch. Its daily active user count surpassed the 20 million mark, reaching 22.15 million, successfully outpacing Doubao's 16.95 million daily active users and accounting for 41.6% of ChatGPT's daily active users. Even more impressive, DeepSeek's monthly active user count (MAU) reached 33.7 million within 21 days of its launch.

Amidst the praise for DeepSeek, Kimi's new model naturally feels somewhat overshadowed. Last year, as the first AI application to break out and achieve massive growth among C-end users, Kimi garnered all the limelight. The new model from Dark Side of the Moon was also highly anticipated within the industry. However, who could have anticipated that DeepSeek would make such a splash by successfully training a model comparable to ChatGPT-4 at an extremely low cost? Compared to this achievement, Kimi's new model's technological advancements appear less groundbreaking.

Considering DeepSeek's growth rate, it seems only a matter of time before it not only surpasses Kimi Intelligent Assistant and ByteDance's Doubao but also rivals ChatGPT globally.

DeepSeek's "surprise attack" on Silicon Valley tech companies has not only rattled OpenAI. If DeepSeek secures the "number one" position in the minds of a vast number of domestic users, the allure of AI applications like Doubao and Kimi Intelligent Assistant, which briefly led the market, will naturally diminish.

On social platforms, many users have already noticed the gap between DeepSeek and other models like Doubao and Kimi Intelligent Assistant. After using DeepSeek, some users have even begun to favor it over other large models. One user bluntly stated, "I use both products, but I only use Kimi k1.5 when DeepSeek doesn't respond. It's too mediocre, while DeepSeek always amazes me, especially in terms of imagination."

While shifts in user preferences and public opinion may not cause products like Kimi Intelligent Assistant to lose their appeal in the short term, once DeepSeek's advantages overwhelmingly surpass them, traffic and capital may also abandon Kimi and its parent company. For Dark Side of the Moon, which is known for its lavish spending, this could spell an even greater crisis.

"Marketing-focused" vs. "Technology-driven"

In October 2023, Kimi made its debut, backed by a large model with 100 billion parameters, supporting input of up to 200,000 characters, making it the global leader in dialogue box capacity for the first time. Five months later, Kimi took another significant step, upgrading to 2 million characters.

Using long text as its main advantage was key to Kimi's prominence among domestic large models. However, Kimi's popularity and large-scale user growth were not fueled by technology but by marketing.

Starting from the 2024 Spring Festival, Kimi targeted Bilibili and launched aggressive advertising campaigns. Users frequently encountered a promotional link named "Kimi AI Intelligent Assistant" on the Bilibili homepage or in the video recommendation area. Benefiting from Bilibili, Kimi's traffic surged, pulling ahead of other applications and laying the foundation for it to become a phenomenal AI application.

Kimi ignited the war of AI advertising, drawing more and more large companies and large model startups into the marketing fray. However, amidst China's lack of breakthroughs in generative AI, this move provided a new path for AI applications to tap into the C-end market. Now, with the emergence and popularity of DeepSeek, it has been proven that domestic large models and applications can conquer the market solely based on technological or cost advantages, posing a certain "threat" to Kimi's "marketing-centric" development path.

Over the past year, Kimi's rapid growth on the C-end has demonstrated the potential of a "Killer App" to the outside world. However, DeepSeek's impressive capabilities undoubtedly position it closer to being a "Killer App" than Kimi.

In fact, there have been doubts within the tech community about Dark Side of the Moon's strategy of large-scale spending and reliance on marketing to sustain user growth.

First, relentless advertising continuously raises the customer acquisition cost of AI applications, causing significant strain for startups with limited financing capabilities. According to APPGrowing's data, since March 2024, Kimi's advertising spending has reached tens of millions of yuan almost every month, peaking at 220 million yuan in October and 200 million yuan in November.

Second, while advertising and burning money have indeed led to a surge in user numbers, user retention remains low. Previously, a set of large model retention data circulated within the industry, revealing that the user retention rate for large model products after 30 days was less than 1%. An industry insider noted that retaining users after buying traffic has long been a consensus in the industry, but the current product form may not be sustainable. "Products like Kimi are essentially tools, and their user duration and retention rates are inherently not very high."

The choice between technology and the market has long been a topic of discussion within the context of domestic large models, with different companies making different choices and emphases. This divergence has, to a certain extent, created gaps and fault lines in the current C-end market for AI applications. Originally, Kimi and Doubao's leadership made the outside world more optimistic about the strategy of competing for the "market" facing users. However, the emergence of DeepSeek is likely to become an industry watershed, making the technology route mainstream, while the effects of marketing will only weaken over time.

Of course, it's not that Kimi has failed to balance technology; it's just that its over-reliance on marketing may hinder its future development.

Have Large Models Ended the Era of "Competing" on Long Text?

Long text is the technical backbone of Kimi, and it is precisely because of Kimi that domestic large model companies have engaged in a "competition" to expand context length. However, this has also diminished Kimi's advantage in long text.

For example, in November 2024, Alibaba Cloud's BaiLian officially launched its new Qwen2.5-Turbo model. The model's standout feature is its support for ultra-long context processing of up to 1 million tokens. According to the latest news, MiniMax has open-sourced its latest basic language model MiniMax-Text-01 and visual multimodal model MiniMax-VL-01. These new models significantly extend the input context window, allowing for the processing of 4 million tokens at a time.

In reality, the technical barriers to long text processing are not high. The reason Kimi could successfully establish its brand with long text as its selling point is mainly due to its effective marketing.

When professional vocabulary related to large models confused the general public, Dark Side of the Moon emphasized and explained "long text," forming a deep impression in users' minds and firmly associating the "label" of long text with Kimi. Even though other large models have surpassed Kimi in long text capabilities, Kimi's popularity last year remained unshaken.

However, technically speaking, this also means that long text capabilities alone are difficult to help Kimi build a sustainable competitive advantage. Globally, the competition among overseas large models for long text capabilities has long ended, as they have found that expanding context to a certain limit has limited impact on the overall improvement of model capabilities. Simply put, after context expansion reaches a sufficient level, improving reasoning abilities requires approaching from other angles.

The enthusiasm for competition around long text in China is also likely to cool down. The reason is simple: DeepSeek's performance in tasks such as mathematics, coding, and natural language reasoning is comparable to that of OpenAI GPT-4, especially its achievements in training costs, which have refreshed the tech community's perceptions. Its existence is likely to change or influence the future focus and main direction of competition among large models.

As one investor noted, the future development of the AI industry depends not only on the improvement of computing power but also on technological innovation and cost-effectiveness. Therefore, the flow of funds may also shift, moving away from investments that purely pursue high-end computing power towards enterprises and projects that place more emphasis on technological innovation and cost control.

Whether it's technological innovation or cost control, Dark Side of the Moon does not hold significant advantages. On one hand, within the same timeframe and with the same benchmark of OpenAI GPT-4, the new Kimi k1.5 model launched by Dark Side of the Moon did not attract as much attention as DeepSeek, highlighting the gap. On the other hand, the high training costs and persistent marketing expenses make Dark Side of the Moon lack credibility in cost control.

Additionally, Dark Side of the Moon still faces an unresolved hidden danger, namely the conflict between investors and entrepreneurs. At the end of last year, Yang Zhilin, the founder of Dark Side of the Moon, and Zhang Yutao, the co-founder and CTO, were arbitrated in Hong Kong by investors from their previous company, Recycle Intelligence. The relevant electronic arbitration application has been submitted to the Hong Kong International Arbitration Centre (HKIAC).

To become a favorite of capital, one must also bear the "sweet burden" that comes with it.

Kimi's popularity opened up the development potential of the C-end market for domestic large models, while DeepSeek's explosion in popularity provides a greater opportunity for domestic large models to lead the world. Although fortunes change, this undoubtedly signifies the robust innovation within China's large model industry.

Of course, in this fiercely competitive race of technology and capabilities, it is too early to determine who will emerge victorious and who will fall behind.

Dao Zong You Li, formerly known as Waidaodao, is a new media outlet focusing on the internet and tech industry. This article is original, and any form of reprinting without retaining the author's relevant information is prohibited.