Image source: Generated by Unbounded AI

The chain reaction about DeepSeek continues.

As the beneficiaries of this round of DeepSeek's direct outbreak, major manufacturers have recently taken out real money and continued to increase their investment in AI.

On February 24, Alibaba CEO Wu Yongming announced that in the next three years, Alibaba will invest more than 380 billion yuan to build cloud and AI hardware infrastructure, with a total amount exceeding the total in the past decade. Tencent even announced the reorganization of its business group, connect to DeepSeek, and starts all in AI.

The big model brought by DeepSeek is a "disaster" for other players in the industry. Data decline, business shrinkage, and even strategy forced to reorganize have become the norm. And the protagonist who was hurt in this "disaster" is undoubtedly Kimi.

According to industry insiders' analysis, domestic big models are being applied on the C-end, experiencing the rise and fall of one another. The popularity of DeepSeek has also caused Kimi and Doubao users to fall from the peak. A typical representative is that Tencent Yuanbao quickly ranks first in Apple's download list with the help of its parent company, and it seems not difficult to surpass Kimi's daily active users.

For Kimi, what is even more difficult is how to turn the mind of the user. Even though there is not much difference between k1.5 and DeepSeek, in users' perception, DeepSeek is not only a traffic, but also an ecological empowerment. Whether it is search engines, mobile phones or cars, they have been connected to DeepSeek, and a big model ecosystem similar to "Android" is gradually taking shape. For Kimi, the road ahead is undoubtedly doomed to be rough.

The famous investor Zhu Xiaohu mentioned in an interview that after DeepSeek became popular, even large factories such as Doubao were not easy to catch up. Once the user's mind was formed, it would be too difficult to carry out market education.

Can Kimi make a comeback in the future? Judging from the company's development path, it is not impossible. As the company that has followed the fastest OpenAI technology, Kimi made several models following OpenAI technology such as Kimi Exploration Edition and k0-math last year. Yang Zhilin himself also said in an interview that the future of big models is not only dependent on reinforcement learning, but also on multimodal capabilities. This seems to be similar to OpenAI.

This means that if OpenAI launches disruptive innovation again, Kimi will still be the fastest-following existence that may appear in China. But whether we can turn users' minds through this requires a question mark.

From the perspective of product and ecological, Kimi undoubtedly faces many difficulties. Internally, the battle between founders and investors seems to be continuing. Externally, in the face of the continuous acceleration of followers, what application solutions can Kimi come up with? Questions such as placement, open source, and even multimodal capabilities are all pouring into Kimi. How to answer them is testing the background of this company.

Kimi failed to have a good year

When metaphysics, stocks, and countless hot searches began to bind to DeepSeek, AI applications were entering an unprecedented period of explosion.

The current leader in discourse power is undoubtedly DeepSeek. Whether it is Liang Wenfeng's every move, or every DeepSeek publishing, or corporate cooperation, or industry popularization, it can easily attract attention.

The audience of AI applications is growing rapidly, but as a former track star, Kimi obviously finds it difficult to get this wave of dividends.

In contrast, during the Spring Festival, Kimi's internal and external problems are gradually amplifying.

For internal reasons, the arbitration case that has been shelved for a long time does not complete the settlement as expected, but instead enters the next process of the arbitration case. According to undercurrent reports, the two parties to the Kimi arbitration case, the old shareholder of Loop Intelligence and Yang Zhilin, have completed the payment at HKIAC (Hong Kong International Arbitration Center) in the end of January and late February respectively, and the party has been completed. The more critical protagonist behind the entire incident, Zhang Yutong, may be filed separately.

According to another investor close to negotiations, after investor Zhu Xiaohu publicly condemned and Yang Zhilin responded to the open letter, Yang Zhilin once came out to negotiate with the old shareholders, expressing his willingness to sell some shares, and a small part of them came from Zhang Yutong's transfer - the implication is that Kimi is not willing to cut off Zhang Yutong.

Obviously, the arbitration between Kimi's founding team and Cyclic Intelligence will continue in the short term. Whether arbitration will affect Kimi's financing and management will become the focus of external concerns.

Besides being subject to arbitration, Kimi's products seem to be under pressure from different levels, and it has become an indisputable fact that the sharp decline in data has been made. During the Spring Festival, it was rumored that Zhang Yutong posted content related to further growth of Kimi users in the circle of friends, but judging from the data disclosed by other institutions, it seems that it does not match it. According to the AI ​​application list released by AIcpb.com, Kimi has fallen to third place in January, and its monthly active users have also dropped from 36 million announced by Yang Zhilin in October last year to 19.43 million.

But for the outside world, the most worrying thing is still the large-scale shift of users' minds. To a certain extent, Kimi has become DeepSeek's "flat replacement" product. Although from the perspective of product, k1.5 has remarkable innovation in long text and multimodal directions, it is gradually submerged in the voice of DeepSeek because it is "out of time", and it can be said that there is no splash. The same case also appeared in the recent "paper crash".

On February 18, DeepSeek and the Dark Side of the Moon were released on the same dayTwo papers were published, the former was submitted by Liang Wenfeng on February 16, and the latter, Yang Zhilin's name rarely appeared in a key position as the author of the paper.

The report shows that unlike DeepSeek, which only published one paper, the Dark Side of the Moon also released relevant code. Moreover, these codes have been actually deployed and verified for a year, and their effectiveness and robustness are guaranteed.

But in terms of external evaluation, the NSA popularity of discussing Liang's technical route is obviously higher than that of Yang Zhilin's MoBA. The reason is that on the one hand, Liang Wenfeng personally took action, which continued the shock of R1, and on the other hand, perhaps people would rather see the "rebels" and recognize his firm attitude.

The internal pressure is under arbitration, and the external competition is suppressed by the decline in user volume. The changes in users' minds will further torture Kimi's previous operating strategy. Whether to continue to invest in large quantities and what to support the next round of financing. Technically, when will the accumulation of Yang Zhilin's investment flow be released will become a question that Kimi has to continue to think about.

The narrative that big companies follow Kimi's old path

DeepSeek exposes is not only a low-cost training plan, but also people's disgust with "pseudo-innovation".

In the launch war that started last year, Kimi was undoubtedly the one who was most injured. According to public information, Kimi's monthly active users were 6.25 million in the report released by QuestMobile in September, while by November, Yang Zhilin claimed that the monthly active users reached 36 million.

appGrowning platform data shows that since March 2024, Kimi's advertising has reached tens of millions of yuan almost every month. The amount of advertising in October even reached 220 million yuan. At that time, Kimi also became a hot search for "burning 100 million in 20 days". In November, Kimi's investment amount reached 200 million yuan again. Based on the information on this platform alone, Kimi, with a total of 36 million users, spent nearly 500 million on advertising. (It is only added from media information, for reference only)

But by January 2025, Kimi's monthly active users fell to about 19 million. Obviously, under the narrative of DeepSeek's super AI application, it is difficult to stop the bleeding of Kimi user loss.

Of course, this part of the loss of users may be related to Kimi's beginning to reflect on the placement and cut the budget. According to reports from the interface, Kimi recently decided to significantly shrink its product placement budget, including suspending the placement of multiple Android channels and cooperation with third-party advertising platforms.

But anyway, Kimi's failure in operational strategy is having a ripple effect. Among them, the most obvious example comes from Tencent's personal outcome.

ADX data shows that before February 5, the amount of materials placed by Kimi, Doubao and Tencent Yuanbao were significantly layered, and there was noSmall gap. Afterwards, Kimi's amount of materials was gradually dropped, and it became the lowest among the three around February 15. It has maintained its daily material volume by not exceeding 1,000 sets.

The supply of bean bread increased to a certain extent after February 5, and then fell sharply, maintaining a similar supply as in late January. Compared with this, the changes in Yuanbao are more obvious.

In the past, the intensity of ingots has never been high, and Tencent is also believed by the outside world to be vague in the field of AI big models and slow in action.

However, on February 15, the intensity of ingot supply increased slightly, and on February 18, it rose sharply. So far, the highest amount of materials released in a single day exceeded 11,000, becoming the first of the three. In contrast, the investment of ingots has increased sharply to 6 days (218-2.23), with a total of 55,000 sets of materials being released, and the month-on-month increase of 12,000 sets of materials being released compared with the previous six days (29-2.14), an increase of 345.1%.

On February 22, Tencent Yuanbao surpassed beanbao and rose to second place in the Apple free list in mainland China. Because Tencent has a wealth of applications, this means that both application linkage and mutual empowerment will have better results. To a certain extent, Yuanbao has followed Kimi's route, and it will be possible to even surpass bean bags in the short term.

Leging behind two positions at once, the giants have followed their old path. Obviously, Kimi not only needs reflection, but may also need a thorough change. The DeepSeek case seems to educate Yang Zhilin to return to technology as the foundation. So from a technical point of view, does Kimi still have hope of turning the tables?

When turning the tables, you can see both OpenAI and DeepSeek

If you want to see Kimi's future, perhaps a clearer route lies in the observation of DeepSeek and OpenAI. Among them, DeepSeek will directly affect Kimi's product rhythm. After all, if the time and ideas of DeepSeek's product launch conference are widening, it means that Kimi will further lose its competitiveness. OpenAI, as the so-called closed source and technological leadership, determines the upper limit of Kimi's route development

According to the news, the report shows that DeepSeek may release r2 before May. According to Microsoft, the company is preparing for the latest model of OpenAI, and GPT-4.5 is expected to be released as soon as next week. Microsoft expects GPT-5 to be released in late May. According to the analysis of industry observers' equality in information, what DeepSeek needs to do is to continue to open source when others continue to close the source. r2 should be targeted with the full version of GPT o3, while DeepSeek V4 should be targeted with at least GPT-4.5. The V4+RL-based model should be targeted with the so-called "GPT-5" in the future”.

From this as the benchmark and rhythm, Kimi's success or failure may occur in the next 3 to 4 months. If DeepSeek and OpenAI have completed a new round of updates to further occupy the minds of users, and at the same time, participants in the DeepSeek camp such as Yuanbao continue to increase their investment efforts, then for Kimi, the remaining value may be the acquisition.

In terms of the technical route, Kimi's leading position compared to DeepSeek lies in its multimodal capability, according to Industry insiders analyzed that DeepSeek may only add multimodal capabilities when it needs at least V4. In comparison, Kimi seems to have more accumulation in multimodal.

In Caixin's interview, unlike many big model companies in the industry regard multimodal capabilities as the next step towards AGI (General Artificial Intelligence) after the language model, Yang Zhilin believes that reinforcement learning is more important. "The next step of big model ability is thinking (reinforcement learning) and interaction (multimodal). Multimodal is necessary, but thinking determines the upper limit of big model capabilities. "He also emphasized that several multimodal models of the Dark Side of the Moon are currently under internal testing.

On December 16, 2024, Kimi released the k1 visual thinking model. According to comments from the evaluation agency, the visual ability of k1 supports complex scenes, including unclear photos, mixed multi-question shots, and handwritten writing interference. Even for purely handwritten questions, k1 can accurately identify and give reasoning processes, far exceeding the coping ability of traditional AI.

Subsequently, on January 15, Kimi The multimodal image understanding model API was released. The new multimodal image understanding model moonshot-v1-vision-preview perfected the multimodal capability of the moonshot-v1 model series.

Then k1.5 appeared, showing more multimodal capabilities. However, these are obviously difficult to widen the gap with DeepSeek.

But it has to be admitted that Kimi has certain competitiveness in the direction of multimodal capability and catch up and replicate OpenAI.

Summary, from various dimensions In other words, Kimi's window of opportunity seems to be getting smaller. The influencing factors should not only refer to the progress of OpenAI, but also fully find the difference with DeepSeek. At the same time, internal contradictions seem to be a variable of the future.

The emergence of DeepSeek has, to some extent, completely rewritten the ecological niche of the big model manufacturers represented by Kimi. What is certain is that if another breakthrough cannot be achieved in technology, it means that the knockout round may occur at any time. As for whether there will be real buyers, everything seems to be still in fog.