After Yang Zhilin made an appearance, Kimi raised another 14 billion in financing.
Investment community sources reveal that Dark Side of the Moon (Kimi) is on the verge of completing a new round of financing worth $2 billion (approximately RMB 14 billion). After this round, its post - investment valuation will exceed $20 billion (approximately RMB 140 billion). This round of financing is led by Meituan Longzhu, with China Mobile, CPE Yuanfeng and others participating, and several existing shareholders have also increased their investments.
So far, Kimi's cumulative financing has exceeded RMB 37.6 billion, making it the startup in the Chinese large - model field with the highest cumulative financing. Its annual ARR revenue has also grown from exceeding $100 million at the beginning of March to over $200 million in April. Subscriptions from paying users and API calls have been the main driving forces.
Just two weeks ago, Kimi released the K2.6 model, which has achieved a breakthrough in long - range coding capabilities. It features a 1T parameter MoE architecture, 32B activation, can code continuously for 13 hours, and its Agent can operate autonomously for 5 days, reclaiming the top spot among global open - source models. Almost at the same time, DeepSeek's first - round financing came to light, and subsequently, DeepSeek V4 was officially unveiled. It is rumored that its valuation is approaching $45 billion.
In this way, Chinese models have taken the top two spots on the global authoritative open - source model list. Some compare this to the "Two Bombs and One Satellite" moment in China's AI field. From DeepSeek to Kimi, Chinese AI companies are transforming from catch - up players to rule - redefiners.
When the Valuation Exceeds $20 Billion, Kimi's Ace Emerges
Kimi is constantly defying expectations.
Late at night on April 20th, Dark Side of the Moon released and open - sourced Kimi K2.6, which can code continuously for 13 hours, write or modify over 4,000 lines of code in one go, and complete the development and optimization of an entire complex system. It is Kimi's most powerful code model to date.
In multiple global authoritative benchmark tests, its performance is on par with or even better than that of top - tier closed - source models such as GPT - 5.4, Claude Opus 4.6, and Gemini 3.1 Pro. It deeply integrates code capabilities with visual understanding and can deliver professional - level web applications with high - end design creativity. This could be a game - changer in terms of efficiency for fields such as front - end development and product prototype design.
The "Agent Cluster" architecture driven by K2.6 has undergone a major upgrade, supporting 300 sub - Agents to complete 4,000 collaborative steps in parallel. Compared with the previous generation K2.5, both the task completion rate and delivery quality have been significantly improved.
Whether it's Attention Residuals, MuonClip, or this Agent Cluster upgrade, Kimi is taking the path of "digging a deep foundation to build a tall building." Its continuous investment in the underlying architecture is now translating into product - level competitiveness.
Four days later, DeepSeek V4 followed. In comparison, both companies have made their own explorations in the adaptation of domestic chips. DeepSeek V4 is the first to be deeply adapted to Huawei Ascend chips, and the inference process will support running on domestic hardware.
Kimi offers another approach. Its latest paper, "Prefill - as - a - Service," proposes a heterogeneous hardware inference framework across data centers, allowing different types of domestic chips to handle the Prefill and Decode stages respectively. In actual tests, the throughput has increased by 54%, and the first - token latency has been reduced by 64%. This provides a practical entry point for domestic chips to truly enter the large - model inference chain.
On the Artificial Analysis intelligent index open - source model list, the top five global open - source models are all Chinese models, among which Kimi K2.6 ranks first in the open - source category. According to several developers, K2.6's programming ability is even better than that of DeepSeek V4.
In other words, among global open - source players, Kimi is now one of the most powerful contenders.
This breakthrough is of far - reaching significance. For a long time, Chinese entrepreneurs have been accustomed to regarding Silicon Valley technology as the standard, as if it were the sole source of technological orthodoxy. However, this narrative is now being rewritten. Chinese AI is carving out its own path along the open - source route - using Chinese chips to run Chinese models and making them open - source to the world.
Data also confirms this trend. According to the monitoring platform OpenRouter, in the global API call volume in the first quarter of this year, the weekly Token call volume of Chinese open - source large models has accounted for over 60%, and Kimi and DeepSeek consistently rank among the top.
From Liang Wenfeng to Yang Zhilin
The last time Yang Zhilin was seen in public was at an entrepreneurs' symposium chaired by the Premier on the afternoon of April 10th. Born in 1992, Yang Zhilin was the youngest participant and the only large - model entrepreneur at the event.
This scene reminds people of more than a year ago: at a similar symposium also chaired by the Premier, another person from Guangdong, Liang Wenfeng, the founder of DeepSeek, sat in that position. At that time, DeepSeek had just become a global sensation, while Kimi, led by Yang Zhilin, was facing growth challenges.
In just one year, the situation has changed significantly. DeepSeek has just released its highly anticipated V4 large model, and Kimi has achieved a remarkable turnaround. With innovation in the underlying architecture as its weapon, Kimi's technological breakthroughs are regarded by Silicon Valley as standards that can be followed.
For a long time, the outside world has always liked to compare Kimi with DeepSeek, and the two companies have had several "tacit" confrontations. For example, when DeepSeek R1 made waves in the AI field, Kimi happened to release K1.5 around the same time. This time, the releases of Kimi K2.6 and Deepseek V4 were less than four days apart.
At similar junctures before, DeepSeek seemed to be in the limelight. However, their subsequent developments have followed different trajectories. DeepSeek has become a phenomenon - level "national AI" with a combination of open - source and low - cost strategies, while Kimi has started from innovating the underlying architecture. Although their styles are different, they have the same goal - to make Chinese AI an indispensable part of the global technological ecosystem.
What's even more remarkable is that there seems to be an atmosphere beyond competition between DeepSeek and Kimi. The MLA attention mechanism of Kimi bears the mark of DeepSeek's early public explorations, and the effectiveness of the key Muon optimizer in DeepSeek V4 was first verified by the Kimi team. In fact, DeepSeek even publicly thanked Kimi in its technical report. This kind of technological linkage is quite rare in the highly competitive large - model field.
Silently, these two top - notch AI teams have formed a rare strategic tacit understanding. One paves the way, and the other builds a bridge, taking turns to lead in the uncharted territory. This spirit of open - source collaboration is perhaps the most hidden and touching romance in the context of China's technological narrative.
Isn't the parallel pursuit of these two technological idealists a microcosm of the explosion of China's AI industry?
In recent years, China has been continuously promoting the "AI +" strategy, and the development of large AI models has entered the fast lane, becoming an important engine for the new round of industrial transformation. Chinese large models have emerged prominently in the global open - source ecosystem, with the cumulative download volume exceeding 10 billion times.
Unconsciously, Chinese companies have begun to reshape the global AI rules. At this year's CES, Huang Renxun demonstrated the performance of the next - generation GPU using DeepSeek and Kimi K2 - Thinking as benchmark models. This is the first time in history.
Beyond technology, the founders of Chinese large - model companies have also become the protagonists in the global AI competition. For example, in March, at the NVIDIA GTC conference, Yang Zhilin was the only Chinese large - model founder invited to give a keynote speech. This is the first time a founder of a Chinese large - model company has occupied this position.
These young Chinese people are changing the tide. As a common consensus in the investment circle goes, the global AI technology has entered the "Chinese time," and Chinese top - notch AI companies are about to be re - evaluated.
"The Shares Are Being Snatched Up"
Investors are voting with their feet.
As it seems that DeepSeek's first - round financing is coming to light. According to the latest report from the Financial Times, DeepSeek's valuation in the first - round financing has reached approximately $45 billion, and large funds are seeking to lead the investment. Just a year ago, Liang Wenfeng clearly stated that he had little interest in commercialization and declined several batches of investors who came to negotiate.
On the other hand, Kimi has already sparked a financing storm in the primary market. In less than half a year this year, Kimi has raised over $3.9 billion, and its latest valuation has more than quadrupled compared to about $4.3 billion in November last year. Now, Kimi, which has been established for only three years, has a cumulative financing of over RMB 37 billion, making it the startup in the large - model field with the most financing.
Looking at the primary market, few companies have been able to raise so much money in such a short period in recent years.
An important turning point seems to be in the making. According to Bloomberg, Dark Side of the Moon is considering an IPO in Hong Kong. It has been less than three months since Yang Zhilin said in an internal letter that the company was not in a hurry to go public in the short term.
This change is closely related to the sharp rise of the AI sector in the Hong Kong stock market. After the listing of companies such as Zhipu and MiniMax, their market values have soared three to five times. This may be a precious window of opportunity. However, Dark Side of the Moon has not made an official response to the IPO rumor.
Even so, beneath the surface, the scramble for Kimi's shares is still intense. In the new round of financing, Meituan Longzhu alone invested over $200 million, and most investors simply couldn't get a share.
Judging from the shareholder structure of this round, the financing is jointly promoted by industry and ecosystem players. Industry investors may become the first choice for Kimi's enterprise - level customers. Institutions such as China Mobile can provide computing power and channel support, while the continuous investment from existing shareholders provides long - term ammunition for this high - intensity technological arms race.
Currently, with the sharp rise of the AI sector, it has become more urgent than ever for investors to "get on board" before the company goes public.
In response, Kimi officially reminds that all financing activities are directly managed by the company, and any so - called "investment shares" circulated through non - official channels pose significant risks.
This situation is thought - provoking.
"One day in AI is like a year in the real world." In an era when everyone is eager to build, expectations and challenges always come hand in hand.
This article is from the WeChat official account "Investment World" (ID: pedaily2012), written by Zhou Jiali and re - published by 36Kr with authorization.