Multiple sources involved in the deals revealed to *Caijing* that the large model company DeepSeek has resumed its second round of financing. This round plans to raise 50 billion yuan, with a pre-money valuation of approximately 500 billion yuan, and aims to complete the signing in late August.
DeepSeek initiated its first round of financing in April this year and completed the closing in June. That round raised 50 billion yuan with a valuation exceeding 350 billion yuan, making it the company with the largest first-round financing scale in the history of Chinese AI large models.
Multiple investors mentioned that DeepSeek's second round of financing had started at least by mid-July, but was suddenly paused at the end of July. At that time, some investors on the negotiation waitlist were informed that plans to sign financing agreements were temporarily on hold.
According to media reports on July 26, one reason for DeepSeek pausing its second round of financing was that DeepSeek's founder, Liang Wenfeng, was dissatisfied with widely circulated online remarks, which mainly revolved around a suspected leaked "meeting transcript for investors."
Caijing sought confirmation from DeepSeek regarding the above market information but received no response as of press time.
The aforementioned deal sources told *Caijing* that DeepSeek and the investors currently at the negotiating table hope this financing round can proceed discreetly after resuming. Currently, some investment institutions that were actively engaging earlier indicated they have not yet received news of DeepSeek resuming financing, and the channel remains in a suspended state.
According to multiple deal sources, the negotiation list for the second round of financing includes both institutions that were "unsuccessful" in the first round and VC (venture capital) and industrial capital that participated in the first round and are considering additional investment. The investors in DeepSeek's first round of financing were: the National Artificial Intelligence Industry Investment Fund, Tencent, CATL and Puquan Capital, NetEase, JD.com, Lichen Capital, IDG Capital, Zhenxingu Investment, and Shixiang Capital.
Since opening for financing, DeepSeek's shares have been in high demand. An investor involved in the deals told *Caijing* that the scale of funds expressing investment interest in the first round around DeepSeek exceeded 1 trillion yuan. Ultimately, 500 billion yuan was raised, meaning there were at least another 500 billion yuan in funds holding intent to enter. These fund providers continued to compete for new shares after the second round opened.
The pre-money valuation for the second round of financing is 500 billion yuan, an increase of approximately 43% compared to the first round. If this round proceeds smoothly, DeepSeek will have raised over 1 trillion yuan across two financing rounds.
DeepSeek's financing scale far exceeds that of its competitors, including Moonshot AI, which has seen heated financing activity recently. These star AI large model companies have significantly accelerated their financing pace this year.
Moonshot AI completed its Series F financing at the end of July with a valuation of approximately $31.5 billion (equivalent to about 2.1 trillion yuan) and immediately initiated a new pre-IPO financing round at a $50 billion valuation after that round closed. According to informed sources, Moonshot AI's current pre-IPO round is expected to complete in August and file for listing within this year.
Alongside financing developments, the model releases from these star companies also stir market sentiment. Model capability and efficiency directly relate to the primary market's valuation of model companies. In an era where model iteration speed is measured in months and leaderboard rankings change weekly, investors hope to secure a ticket on the ship that first reaches AGI (Artificial General Intelligence) amidst uncertainty. A deal source believes, "Valuing large models is essentially an option, not a financial model based on cash flow."
According to previous *Caijing* reports, the Series F financing initiated by Moonshot AI in June initially saw lukewarm interest, with many channels receiving allocations "unsold for several days." At that time, even secondary market transactions at a $18 billion valuation were circulating. However, after the release of its new flagship model Kimi K3 and its global acclaim, its shares suddenly became highly sought-after, with some sellers of secondary shares even raising prices.
An investor close to DeepSeek told *Caijing* in late July that the pressure on DeepSeek to train new models recently has not been small. The official version of DeepSeek-V4, originally scheduled for release in mid-July, only launched for public beta in the latest official API documentation log on July 31. However, only V4-Flash debuted; the official version of V4-Pro has not yet been released, with a specific public beta time still unannounced.
According to official documentation, DeepSeek-V4-Flash made no adjustments to model architecture or parameter size. Instead, it strengthened post-training on the original foundation, significantly enhancing its Agent capabilities, with multiple benchmark tests far exceeding the V4-Flash preview version released in April.
According to the latest Intelligence Index released by the independent third-party AI model evaluation platform Artificial Analysis, the intelligence level of DeepSeek-V4-Flash official version reached 50 points, ranking second domestically only to Zhipu's GLM-5.2 at 51 points, and higher than the average level of similar models. DeepSeek continues its model strategy of "low cost, high cost-effectiveness." The price for V4-Flash output per million Tokens is $0.28, lower than Zhipu GLM-5.2's pricing of $4.29 per million Tokens output.
With rising demand for inference computing power and continuously pushing costs higher, many large model companies are pursuing extreme inference efficiency. Domestically, companies following a high cost-effectiveness route also include Tencent's Hunyuan. Recently, the American AI model company OpenAI also announced price reductions for its GPT-5.6 series model API calls, with the entry-level model Luna reduced by 80%. Before the release of the DeepSeek-V4-Flash official version, GPT-5.6's cost-effectiveness was even higher than DeepSeek's for the same model capability. However, this cost-effectiveness advantage was quickly eliminated.
On the latest OpenRouter chart, an AI model aggregation platform for global developers, DeepSeek-V4-Flash official version ranked first this week with 7.1 trillion Tokens consumed.
