Kimi, the Resurgence of Empress Xi

07/18 2025 474

By She Zongming

For many "Huanxue scholars," the zenith of "Empresses in the Palace" lies in the triumphant return of Empress Xi.

After enduring the ordeal of Chunyuan's discarded garments and her father's exile, Zhen Huan lost her power and willingly departed the palace.

Upon her arrival at the Sweet Dew Temple, many believed her days of glory had ended.

After all, there was no precedent for a disgraced concubine to reclaim her place in the palace.

Yet Zhen Huan remains Zhen Huan. Clad in simple attire, she bided her time in the mountain temple, awaiting the opportune moment.

As many are aware, Zhen Huan made her second return to the palace, steering the course of her destiny.

Similarly, The Dark Side of the Moon, which once shone brightly but then retreated into hibernation, seems to have followed Zhen Huan's script.

With the recent release of Kimi K2, The Dark Side of the Moon has staged a remarkable comeback. The overseas technology forum buzzed with discussions, and Huang Renxun gave it a thumbs-up, while Musk commented, solidifying The Dark Side of the Moon's status as the "coolest kid on the AI megamodel block" once again.

At the end of July, AI startups such as Jieyue Xingchen and Zhipu AI will also unveil their self-developed new generation of foundational megamodels.

They all aspire to prove that they are still major players in the game.

01

At the dawn of 2023, following ChatGPT's meteoric rise, domestic tech giants and AI startups swiftly entered the fray.

Amidst the intensifying competition, the "AI Six Little Tigers" swiftly emerged as industry leaders. Throughout 2023, the cumulative funding of the "AI Six Little Tigers" accounted for over 50% of domestic megamodel vendors.

Upon closer inspection, Zhipu AI's B-end ecosystem, Jieyue Xingchen's multimodal terminal Agent, MiniMax's MoE mixed expert model—each of these star startups excels in its unique domain.

Furthermore, in March 2024, The Dark Side of the Moon ignited a fervor for concept stocks due to its long-text capabilities. Akin to the Nine Yin Manual, these capabilities propelled Kimi to deity-like status in a single battle. Among a constellation of megamodels renowned for their massive parameter scales, Kimi swiftly became a sensation.

Two months later, Tencent's investment, following Sequoia China, Meituan, and Alibaba, propelled its valuation to $3.3 billion, making The Dark Side of the Moon a topic of conversation for some time.

At this juncture, it's akin to the progress bar of "Empresses in the Palace" reaching episode 8: In this pivotal episode, Zhen Huan received the "spiced chamber favor."

Had subsequent events not unfolded, the narrative surrounding Kimi would have continued along the lines of "a brilliant young leader guiding a star startup to the pinnacle," portraying a picture of AI startups where "Zhang Hua was admitted to Peking University, Li Ping entered a mid-level technical school... We all have promising futures."

However, reality seldom adheres to such scenarios.

02

Entering the second half of 2024, the industry reached a turning point, and megamodel vendors entered a critical testing phase.

Among the "AI Six Little Tigers," Baichuan Intelligence, which had previously vowed to benchmark its underlying model against OpenAI, ceased training ultra-large general models from 2024 onwards, instead focusing on the medical field. ZeroOne, which claimed to be "building a brand-new AI 2.0 platform," laid off its entire pre-training algorithm team at the end of 2024, with the Infra team being taken over by Alibaba Cloud.

Zhipu and Jieyue Xingchen, which rely on government resources, are increasingly focusing on the government and enterprise market. MiniMax is concentrating on multimodality and has also partially ventured into overseas markets.

The Dark Side of the Moon, which remains enthralled with AGI, appears to have shifted its technical focus entirely to "reinforcement learning" in the latter half of the year, successively releasing mathematical models and visual thinking models, while adopting an aggressive product promotion strategy to ensure Kimi doesn't lag behind in the competition with products like Doubao, which also target the C-end market.

A more significant shift was instigated by DeepSeek. The emergence of DeepSeek R1 earlier this year altered the already established landscape of the megamodel industry. With its "new architecture + low cost + open-source ecosystem," DeepSeek successfully ascended to the pinnacle.

The market share of megamodel products exhibits a power-law distribution effect, and the public opinion field has a frequency domain shielding effect. Under the spotlight of DeepSeek's "national-level achievement," the "AI Six Little Tigers," which once firmly held the C-position, have seemingly become mere foils.

Many have withdrawn from the "ultra-large base model competition," the iteration speed of megamodels has slowed down, and only one company has announced funding news since the beginning of the year... Numerous individuals have pieced together evidence suggesting the "disappointment" of the "AI Six Little Tigers."

The Dark Side of the Moon was eclipsed by comparisons to "Liang Yu." The Kimi 1.5 reasoning model was released on the same day as DeepSeek R1, but its popularity was utterly overshadowed. Thereafter, events unfolded in the direction of "Huanhuan is like Qing but not Qing": After DeepSeek and Kimi were juxtaposed, the tendency of public opinion to "praise one and suppress the other" morphed into criticism directed at both.

Subsequently, The Dark Side of the Moon began to reflect and halted its aggressive promotion strategy.

Consequently, user retention faced challenges. Even when searching for Kimi on search engines and app stores, other products that continued to invest in advertising emerged prominently.

Looking back at the first half of 2025, abroad, OpenAI's GPT series, Anthropic's Claude, Google's Gemini, and Musk's xAI Grok took turns to make their mark; domestically, there were also constant advancements.

However, the tone of public discourse regarding the "AI Six Little Tigers" tended to be "gloomy."

03

Yet, with the BGM of "My stubbornness and I, holding hands tightly and never letting go," some of the "Six Little Tigers" continued to make their presence felt.

MiniMax released five products in five consecutive days, and Zhipu, on the cusp of an IPO, also launched an open-source visual understanding model capable of watching "Scottish Premiership" matches.

The Dark Side of the Moon even pulled off a major coup. Kimi is akin to Zhen Huan in the AI realm. When others presumed she had fallen out of favor, she proved many wrong by declaring, "A stronger Zhen Huan has returned."

Given C-end users' limited perception of technological innovations with high professional thresholds, many may not fully grasp the significance of Kimi K2 achieving open-source SOTA (state-of-the-art) performance in tests involving code, agents, tool invocations, etc.

Against the backdrop of the public's numbness to shock headlines like "strongest" and "explosive" in AI narratives, many may also fail to comprehend the implications of Kimi K2 surpassing xAI in Token consumption within just two days of its launch on the OpenRouter platform, topping the global API call growth chart, and ranking highest in API usage among global open-source models on platforms such as Cline, Roo Code, and Kilo Code.

However, Huang Renxun's statement in an interview that "DeepSeek R1, Qwen, and Kimi are the best open-source models in the world," and a report by CNBC stating that Kimi K2 outperforms Claude Opus 4, surpasses OpenAI's GPT-4.1, and boasts lower costs in benchmark tests, have already elucidated much.

The Nature magazine website even hailed the release of Kimi K2 as "another DeepSeek moment for the world," imbued with a sense of "Make Kimi Great Again."

Moving forward, there are still a few companies poised to play their cards. Judging by the response that Kimi K2 has elicited in the global tech community, many are already brimming with anticipation: How many more surprises do you have up your sleeve that I'm unaware of?

04

Now, it appears that the Six Little Dragons, who were ensnared in a narrative of solitude a few months ago, have used their actions to affirm two pivotal points:

1. AI competition is a marathon, a long-term endeavor.

What determines a company's trajectory is not whether it will fall into a trough but whether it can rebound after doing so.

2. Product strength is the ultimate marketing, and a superior product speaks for itself.

Following the launch of Kimi K2, Aravind, CEO of the unicorn Perplexity, stated on social media that based on the outstanding performance of the Kimi K2 model, the company would utilize K2 for post-training. The last Chinese model employed by the company for technical training was DeepSeek R1. Thomas, co-founder of Hugging Face, the world's largest open-source AI community, remarked that the K2 model, which consistently breaks through limitations and challenges closed-source models, is incredible.

These indicators suggest that the window of opportunity for AI startups is far from closed.

However, you must first "build momentum"—that is, accumulate strength.

05

Intriguingly, on July 14, Nathan Lambert, a renowned AI researcher in the United States, penned an article titled "Kimi K2 and When DeepSeek Moments Become the Norm."

In the article, he bluntly asserted that Kimi K2 is already competitive with leading cutting-edge models in the United States, "China is constantly approaching or even reaching the absolute frontier of model performance, and we are on the path to 'failure,' with the West further lagging behind in open-source models."

Azeem Azhar, a Silicon Valley AI entrepreneur and author of the renowned newsletter Exponential View, wrote an article titled "Kimi K2 Should Worry Silicon Valley."

He stated, "Sputnik 1 proved that the Soviet Union could breach Earth's gravity and shatter Western technological dominance. Four years later, the Vostok 1 spacecraft carried Yuri Gagarin around the Earth, confirming that Sputnik was no fluke. In today's realm of artificial intelligence, DeepSeek played the role of Sputnik, and now, artificial intelligence has ushered in its Vostok 1 moment."

These words subtly convey a message: DeepSeek and Kimi are the twin stars of Chinese AI shining brightly in the world, and they have imparted another lesson to the United States.

From the ascendancy of DeepSeek R1 to Kimi K2, it is no coincidence: Both DeepSeek and The Dark Side of the Moon have been consistently building momentum and have met several times at the pinnacle.

In February of this year, Liang Wenfeng and Yang Zhilin published papers on roughly the same day, both delving into the core attention mechanism of the Transformer architecture. In April, The Dark Side of the Moon launched Kimina-Prover, a large model for formal theorem proving, and DeepSeek swiftly followed up with DeepSeek-Prover-V2, a dedicated model for mathematical theorem proving, surpassing Kimi. And in July, Kimina-Prover-7B-Distil continued to update, achieving a reversal.

To some extent, from the DeepSeek moment to the current Kimi moment, they have provided a roadmap for the ascent of Chinese AI startups: The paramount importance lies in building momentum in original innovation.

In "Empresses in the Palace," Zhen Huan declared before her return to the palace, "Rather than admiring from afar, it's better to be that person yourself." "That person" refers to someone who can master their own destiny.

Yang Zhilin, an avid fan of the Pink Floyd album "The Dark Side of the Moon" who has heard them sing "Don't sit down, it's time to dig to another... Only by riding the wave can you find balance amidst the giant waves," has demonstrated with Kimi K2 that he is striving to become "that person."

Next, more AI startup leaders need to prove: They too can become "that person."

Chinese AI startups will inevitably diverge and chart different courses, but it is certain that Kimi has "returned as Empress Xi," and other little dragons and tigers are also continuing to "showcase their prowess." This grand saga still harbors many thrilling moments ahead.

Solemnly declare: the copyright of this article belongs to the original author. The reprinted article is only for the purpose of spreading more information. If the author's information is marked incorrectly, please contact us immediately to modify or delete it. Thank you.