DeepSeek has restarted its second funding round, Caijing reported, citing multiple trading sources. The large-model company is seeking RMB 50 billion at a pre-money valuation of about RMB 500 billion, with signing planned for late August.
According to the report, DeepSeek opened its first financing round in April and completed the transaction in June. That round raised RMB 50 billion at a valuation above RMB 350 billion, making it the largest first-round financing on record among China’s AI large-model companies.
Several investors told Caijing that the second round had begun by at least mid-July, but was suddenly paused at the end of July. At that point, some investors on the standby list for negotiations were told that plans to sign financing agreements had been put on hold.
A July 26 media report said one reason for the pause was founder Liang Wenfeng’s dissatisfaction with widely circulated online claims centered on what was described as a leaked “meeting transcript for investors.” Caijing said it sought comment from DeepSeek on those market reports but had not received a response by the time of publication.
One of the trading sources told Caijing that DeepSeek and the investors still at the table want the restarted financing process to proceed quietly. Some institutions that had previously been active in discussions said they had not yet been informed that the round had resumed, and access channels were still on hold.
People familiar with the talks said the second-round discussion list includes both standby institutions that missed out in the first round and existing venture capital and industrial investors considering follow-on commitments. Investors in DeepSeek’s first round were the National AI Industry Investment Fund, Tencent, CATL and Puquan Capital, NetEase, JD.com, LISI Capital, IDG Capital, Zhengxingu Investment, and Shixiang Capital.
Since opening up to outside capital, DeepSeek’s allocation has remained oversubscribed, the report said. One investment participant told Caijing that more than RMB 100 billion had expressed interest in the first round, while only RMB 50 billion was ultimately raised, leaving at least RMB 50 billion still looking for a way in. Those investors moved to compete for new allocation once the second round opened.
At a reported pre-money valuation of RMB 500 billion, the second round would represent an increase of about 43% from the first round. If completed as planned, DeepSeek would raise more than RMB 100 billion across the two financings.
Caijing said DeepSeek’s fundraising scale is well ahead of rivals, including Moonshot AI, another high-profile large-model company that has also accelerated financing activity this year.
Moonshot AI completed its Series F round at the end of July at a valuation of about $31.5 billion, or roughly RMB 210 billion, the report said. It then immediately launched a new Pre-IPO round at a $50 billion valuation. Caijing also cited a person familiar with the matter as saying the Pre-IPO round is expected to close in August, with a listing filing planned within the year.
Model launches are drawing as much attention as fundraising. Caijing quoted one trading source as saying, “The pricing of large models is essentially an option, rather than a financial model based on cash flow.”
The outlet also revisited its earlier reporting on Moonshot AI’s Series F round, which began in June with relatively muted demand. It said some channels that obtained allocation “couldn’t sell it after several days,” and old-share deals at a $18 billion valuation were circulating in the market. After the release of the new flagship model Kimi K3 and the attention it drew globally, demand for the company’s shares picked up sharply, and some sellers of old shares raised their asking prices.
A person close to DeepSeek told Caijing in late July that the company was under pressure in training its new model. The formal version of DeepSeek-V4, originally scheduled for release in mid-July, did not enter public beta until July 31 through the latest log in its official API documentation. Even then, only V4-Flash appeared. The formal version of V4-Pro has not been released, and no specific public beta date has been announced.
According to the official documentation, DeepSeek-V4-Flash did not change its model architecture or parameter size. Instead, it strengthened post-training on top of the existing foundation, significantly improving its agent capabilities. The report said multiple benchmark results were far ahead of the V4-Flash preview released in April.
Artificial Analysis’ latest Intelligence Index gave the formal version of DeepSeek-V4-Flash a score of 50, placing it second in China behind Zhipu GLM-5.2 at 51 and above the average for comparable models. On pricing, the report said DeepSeek has stuck with a low-cost, high-value strategy: V4-Flash charges $0.28 per million output tokens, versus $4.29 per million output tokens for GLM-5.2.
Caijing added that as inference demand rises and costs keep climbing, many large-model companies are pushing for maximum inference efficiency. Tencent Hunyuan was named as another domestic model following a high-value route. In the US, OpenAI has also cut API prices for the GPT-5.6 series, with the entry-level Luna model reduced by 80%. Before the formal release of DeepSeek-V4-Flash, GPT-5.6 had offered better value at the same capability level, according to the report, but that edge quickly disappeared.
On OpenRouter’s latest ranking for global developers, the formal version of DeepSeek-V4-Flash ranked first this week by token consumption at 7.1 trillion.