Source: LatePost
By Wei Bing
Edited by Cheng Manqi
LatePost has exclusively learned that Moonshot AI (Kimi) submitted its A1 application to the Hong Kong Stock Exchange (HKEX) on a confidential basis this week, formally initiating its IPO process in Hong Kong. The A1 form is one of the official documents required for companies applying for a listing on the HKEX.
In response to the aforementioned reports, a representative from Kimi stated: “We do not comment on market rumors, and there is currently no information to disclose.”
Meanwhile, we understand that Kimi is advancing a new round of financing with a pre-money valuation of $50 billion. This is likely the final funding round before Kimi’s IPO.
It is widely believed that DeepSeek, another leading Chinese large-model venture yet to go public, may also list in the first half of next year.
At the end of 2025, Kimi founder Yang Zhilin had stated that there was no urgent need for an IPO in the short term. Since the beginning of this year, however, Kimi has significantly accelerated its model iterations, commercialization efforts, and capital activities: K2.5 was launched in January, K2.6 in April, and K3 in July, maintaining an update cycle of approximately every three months. During the same period,as previously reported by us, Kimi’s Annual Recurring Revenue (ARR) surpassed $100 million in early March,and the Shanghai STAR Market Daily subsequently reported that this figure had exceeded $300 million by mid-June.
Valuations in the primary market have also risen rapidly. At the end of 2025, Kimi’s valuation was approximately $4.3 billion; by May of this year, it had exceeded $20 billion. Following the release of K3, Kimi completed another financing round around July, reaching a post-money valuation of $35 billion—an eightfold increase in just over six months. Currently, the company is undertaking its latest round of financing with a pre-money valuation of approximately $50 billion.
Unlike traditional software companies, model companies face high marginal costs for additional revenue. More powerful models drive increased usage but also rapidly consume greater computational resources. After the launch of K3, Moonshot AI temporarily suspended subscriptions for new users to prioritize computational capacity for existing paying customers. To support the next round of model iteration and commercial growth, the company requires continuous financing and expansion of its computational infrastructure.
As call volumes and revenues grow rapidly, competition among model companies is no longer limited to model capabilities. Over the past few months,$Z.AI (02513.HK)$、 $MINIMAX-W (00100.HK)$ revenue data has been disclosed successively. ARR, API calls, and enterprise clients have become new operational metrics.
On July 18, the day after the release of K3, Zhang Yutong, President of Kimi, revealed on social media that enterprise ARR had multiplied several times following the K3 launch, marking the largest single-day increase in history.
According to a report by The Information in late August, DeepSeek's actual revenue for the first seven months of this year was approximately RMB 475 million, about ten times its full-year revenue for 2025.
According to financial reports and earnings conference calls, Zhipu AI generated revenue of RMB 954 million in the first half of the year, with ARR reaching USD 1.6 billion in August. MiniMax recorded revenue of USD 117 million (approximately RMB 786 million) in the first half, with ARR exceeding USD 800 million in August.
ARR generally has two meanings:
First, Annual Recurring Revenue, which has primarily served as an operational metric for evaluating SaaS business models. Recurring revenue traditionally refers to software subscription income, characterized by relatively stable monthly amounts, allowing for fairly accurate projections of future revenue changes based on metrics such as renewal rates and upsell rates.
Second, Annual Run Rate, which projects annual revenue based on recent performance. This revenue figure may include subscription fees, API usage charges, and project-based income from customized services.
Companies such as Zhipu AI and MiniMax have not clearly specified which definition of ARR applies in various contexts. In contrast, the ARR figures released by Anthropic this year refer to Annual Run Rate.
The next steps for investment among these companies are also becoming increasingly similar. Based on the latest iteration directions of their models, Zhipu AI is focusing on coding, agents, and long-horizon tasks as key model capabilities. MiniMax aims to further apply its models to coding and enterprise productivity scenarios. DeepSeek's latest model continues to strengthen capabilities in coding, agents, and complex tasks, and recently launched the open-source agent framework DeepSeek Harness, enabling developers to flexibly integrate different models, plugins, and skills to handle more complex and personalized tasks.
The next version of Kimi’s model, K3.1, is also poised for imminent release. In late July, leaks related to K3.1 emerged on social media platforms, describing it as a minor iterative update to K3, primarily optimizing inference speed and agent capabilities. As of now, K3.1 has not been officially released.
K3 has demonstrated the technical prowess of Kimi. Following its entry into the secondary market, the company, along with other AI firms, must continue to address how these technological capabilities can ultimately be translated into greater revenue and profits, thereby delivering enhanced value to users and clients.
Editor/Rocky