It has just been revealed that Moonshot AI has raised another 23.7 billion yuan in financing.

According to a report by AI Technology Review on July 29, citing sources familiar with the matter from Bloomberg, Beijing-based large model unicorn Moonshot AI has raised more than expected 35 billion USD (approximately 236.77 billion RMB) in its latest financing round, with a post-money valuation reaching 350 billion USD (approximately 2367.71 billion RMB).

Insiders also revealed that Moonshot AI originally planned to raise 10 to 20 billion USD in this round of financing. Calculated based on this figure, the actual amount raised in this round is 75% oversubscribed compared to the original upper limit, reaching 3.5 times the original lower limit. At present, Moonshot AI has begun to contact potential investors and plans to launch Pre-IPO financing with a pre-money valuation of 500 billion USD (approximately 3383.25 billion RMB). Moonshot AI hopes to complete this round of financing within this year and then list in Hong Kong as soon as possible.

AI Technology Review has reached out to Moonshot AI for comment on the above news, and the company has no comment for the time being.

Moonshot AI has made frequent moves in the capital market recently. According to Qichacha, an enterprise information query platform, on May 7 this year, Moonshot AI completed a financing led by Meituan, with a post-money valuation of 20 billion USD; on June 30 this year, Moonshot AI secured its Series E financing, with investors including state-owned capital from Shanghai, Shenzhen and other regions, as well as investment institutions such as IDG Capital, and the specific financing amount and valuation have not been disclosed yet.

If the post-money valuation of 350 billion USD is true, Moonshot AI's valuation will rise from 4.3 billion USD in December last year to more than 8 times the original level in about 7 months, with a growth rate of about 714%.

In terms of financing scale, the 35 billion USD raised by Moonshot AI in this round is second only to the approximately 51 billion USD financing completed by DeepSeek in June this year, setting the second largest record for a single round of financing by a Chinese large model company.

Calculated based on the closing price of Hong Kong stocks on July 29, Moonshot AI's valuation has exceeded the market value of MiniMax at 74.217 billion HKD (approximately 64.121 billion RMB), roughly 57.07% of Zhipu AI's market value of 4809.89 billion HKD (approximately 4149.97 billion RMB).

Moonshot AI's financing history (Source: Qichacha)

The huge popularity of Moonshot AI's new flagship model Kimi K3 is an important background for the over-subscription of this round of financing.

On July 16, Moonshot AI released its new-generation model Kimi K3, which has a total parameter scale of 2.8 trillion, is built on Moonshot AI's self-developed KDA hybrid linear attention mechanism (Kimi Delta Attention) and Attention Residuals technology, natively supports visual understanding, has a 1 million token context window, and is currently the open-source model with the largest parameter scale in the world.

According to the architecture information revealed on the official blog, K3 adopts the Stable LatentMoE framework to further increase the sparsity of MoE, with 16 experts activated out of 896 for each token; combined with the optimization of training and data recipes, the overall scaling efficiency of K3 is about 2.5 times higher than that of K2.

The complete model weights and technical report of K3 have been open-sourced on July 27, along with inference and training toolchains and technical reports such as MoonEP, FlashKDA, and AgentEnv.

The release of this model has caused a great response in the global AI community. Elon Musk left a comment under K3's evaluation post, saying that this model is impressive, and a few days later he listed K3 as the benchmark model for Grok 4.5. Professor Stoica from the University of California, Berkeley also made a judgment: Previously, the industry generally believed that China's open-source models were 6 to 9 months behind the most advanced level, and now this gap may have narrowed to 2 to 3 months.

Elon Musk lists Kimi K3 as the benchmark for Grok 4.5 (Source: X Platform)

The technical reputation of K3 quickly translated into traffic pressure. Three days after the release of K3, late at night on July 19, the Kimi team of Moonshot AI released the "Statement on Computing Power Shortage and Suspension of Member Registration", saying that since the launch of K3, "in the past 48 hours, user requests have far exceeded our estimates and are approaching the carrying limit of the existing cluster".

Moonshot AI's statement on computing power shortage and suspension of member registration (Source: Moonshot AI)

The disposal plan given in the announcement is: from now on, suspend new C-end user subscriptions, put all existing computing power into serving subscribed users, and ensure that all rights and interests of subscribed users are not affected; at the same time, fully advance the expansion of computing power, and gradually open up more subscription quotas after new computing power is put in place one after another, until the normal subscription is fully restored.

Conclusion: Hot Money Flocks to Large Models, Financing, Listing and Hoarding Computing Power

This round of over-subscription of Moonshot AI reflects the high financing heat in China's large model track at present. After Zhipu AI and MiniMax successively listed on the Hong Kong Stock Exchange, their market values have increased several times. The financing scale and valuation of companies such as DeepSeek and Moonshot AI are also growing rapidly, and capital is accelerating to concentrate in this field.

Next, the focus of the industry will turn to the listing rhythm. Companies including Moonshot AI, DeepSeek, and 01.AI have all reported news of Hong Kong stock IPOs. Large-scale financing is expected to help these enterprises reserve more computing power and talent resources to support the R&D, training and inference of the next generation of models.

This article is from the WeChat official account "AI Technology Review" (ID: zhidxbc), written by Chen Junda, edited by Li Shuiqing, and published with authorization by 36Kr.