In 20 days, the valuation soared by 3.4 billion yuan. A Beijing AI unicorn announced new financing.

On January 20th, according to foreign media CNBC, two sources familiar with the matter revealed that Beijing-based large model unicorn Dark Side of the Moon is conducting a new round of financing. This round of financing values the company at $4.8 billion (approximately RMB 33.413 billion), while its valuation was only $4.3 billion (approximately RMB 29.932 billion) in the Series C financing announced just 20 days ago.

In addition, the above-mentioned sources revealed that due to strong market demand, this round of financing is likely to be completed soon. As market interest in Chinese AI IPO candidates surges, Dark Side of the Moon may see further room for its valuation to rise in subsequent financing rounds.

TechNode contacted Dark Side of the Moon for verification on the above content, and the company said it has no more information to share for now.

In an internal letter exposed on December 31, 2025, Yang Zhilin, the founder and CEO of Dark Side of the Moon, revealed the company's current financial situation. Dark Side of the Moon recently completed a $500 million (approximately RMB 3.48 billion) Series C financing, significantly oversubscribed, and currently holds over RMB 10 billion in cash.

Yang Zhilin said that compared with the secondary market, Dark Side of the Moon believes it can raise more funds from the primary market. The company is not in a hurry to go public in the short term but plans to use listing as a means to accelerate AGI in the future and will act at the right time.

Technically, Yang Zhilin shared many details at the AGI Next Summit on January 10th. Currently, the pre-training strategy of the Kimi model focuses on two main lines: Token efficiency and long context ability.

In terms of improving token efficiency, Dark Side of the Moon proposed an improved version of the Muon optimizer last year, achieving approximately a twofold increase in Token efficiency. It also solved the Logits explosion problem through QK-Clip, supporting the stable training of trillion-parameter models. The improvement in Token efficiency means that Dark Side of the Moon can obtain more intelligence from the same amount of data.

In terms of long context ability, Dark Side of the Moon proposed a new linear attention mechanism called Kimi-Linear, which can significantly improve the inference speed while ensuring the effect, providing a practical path for ultra-long context tasks.

On the model and product side, Dark Side of the Moon open-sourced the trillion-parameter model Kimi K2 and its corresponding inference model Kimi K2 Thinking last year and also launched the OK Computer Agent mode.

Statistics from the large model calling platform OpenRouter show that the call volume of the Kimi K2 series models has occupied a certain market share after its release.

From November 2024 to November 2025, on the OpenRouter platform, the call volume of Dark Side of the Moon's models was 0.92 trillion tokens, ranking ninth globally among open-source model players, slightly less than that of Zhipu (1.18 trillion tokens) and MiniMax (1.26 trillion tokens). OpenRouter data mainly reflects the usage in the overseas market.

In this year's CES speech by Jensen Huang, the founder and CEO of NVIDIA, Kimi K2 Thinking replaced the commonly seen DeepSeek-R1 at previous NVIDIA press conferences and was used to demonstrate the performance of NVIDIA's Rubin NVL72 system.

Regarding the evolution direction of the next-generation model Kimi K3, Yang Zhilin shared at the AGI Next Summit that Dark Side of the Moon will further scale on the basis of K2. However, scaling is not just about increasing computing power but also a comprehensive evolution of architecture, data, and "taste."

He especially emphasized that creating a model is essentially creating a worldview, and the model also needs to have better taste and aesthetics. He believes that the intelligence generated by the model is essentially a Non-Fungible Token (NFT), which carries taste, aesthetics, and values. In the dimension of intelligence, taste has an extremely broad exploration space, and this space expands exponentially, which is a very important goal for Dark Side of the Moon in the future.

Conclusion: Leading Chinese large model companies are ready for the next stage

Since the beginning of 2026, many leading Chinese large model companies have achieved great success in the capital market. The stock prices of Zhipu and MiniMax, which went public one after another, have soared. MiniMax has even become the first domestic AI large model company with a market value exceeding RMB 100 billion.

With its valuation continuously rising in a short period and abundant cash reserves, Dark Side of the Moon may have the confidence to continue to increase its investment in the future. Previously, Yang Zhilin revealed that the Series C financing would be used to more aggressively expand the number of graphics cards to accelerate the training and R & D of the K3 model.

It is worth continuing to pay attention to whether Dark Side of the Moon can occupy a more stable position in the Chinese and even global large model landscape.

This article is from the WeChat official account "TechNode" (ID: zhidxcom), author: Chen Junda, editor: Li Shuiqing. It is published by 36Kr with authorization.