Liang Wenfeng is opening the company's doors to strategic investment for the first time. On May 8th, news emerged that DeepSeek, under his leadership, is nearing the completion of its first funding round, with a post-money valuation reaching $50 billion, approximately 340 billion RMB. As the founder of Hangzhou DeepSeek, Liang Wenfeng holds an ultimate beneficial ownership stake of 84%. It is reported that his personal investment could be as high as 20 billion RMB, accounting for 40% of the total funds raised.

Just before the fundraising effort, he had just "handed in his homework." On April 24th, DeepSeek-V4 was launched, a full 484 days after the release of the previous major version, V3. "DeepSeek has lost too much talent, and access to high-end computing power is limited." A veteran AI industry practitioner indicated that this is the reason for the shift in its financing stance and the delay in the new version's release. Stepping down from the ivory tower of idealism, Liang Wenfeng is seeking powerful allies to confront the practical challenges beyond technological innovation.

An Anchor to Retain Talent DeepSeek is not short on capital. In 2025, the average return rate of its parent company, Huanfang Quant, reached 56.6%. Based on an assets under management (AUM) scale of 70 billion RMB, this corresponds to roughly 40 billion RMB in profit. Following industry-standard profit-sharing arrangements, Huanfang Quant could take away 7 to 8 billion RMB. With Huanfang Quant as a financial backer, providing funding for DeepSeek's R&D, the company has a substantial financial safety net. "The problem we face has never been money," Liang Wenfeng once stated. However, employees have personal financial considerations, and DeepSeek's brightest minds are leaving. A review of DeepSeek V4's 58-page technical report reveals an author list at the end where Liang Wenfeng is listed alongside nearly 300 other researchers and engineers. Ten names marked with asterisks stand out, annotated as "former employees." One of the recent departures among technical backbones is Guo Daya, who was deeply involved in the development of hit models like V3 and R1, and joined ByteDance's Seed unit with a rumored annual salary approaching 100 million RMB. ByteDance denied this claim, adding a teaser: "It's not impossible that some Seed technical personnel could see returns reaching several hundred million RMB after four years." In reality, Liang Wenfeng offers competitive salaries. A job applicant revealed that during a 2024 interview for a DeepSeek deep learning researcher position, HR indicated a potential salary of 1.5 million RMB. This "tech geek" boss also provides ample freedom. One detail is that DeepSeek employees typically leave work at 6 PM, with no morning clock-in and no KPIs. This stems from Liang Wenfeng's belief that a person's capacity for high-quality output in a day rarely exceeds 6-8 hours, and that innovation requires minimal intervention and management. In the AI talent war, Liang Wenfeng needs to provide his team with more security, particularly regarding the value of employee stock options. With external financing, DeepSeek will have a public valuation, providing a pricing anchor for stock options. This is a prerequisite for retaining talent. Liang Wenfeng has changed his stance and is currently in talks with institutions such as the National Integrated Circuit Industry Investment Fund and Tencent. Informed sources indicate that this funding round will be used to enhance computing power and improve employee benefits to cope with intense competition.

The Computing Power Battle Beyond the talent issue, Liang Wenfeng faces a tough battle over computing power. As of May 9th, DeepSeek had 36 open positions urgently needing to be filled. On April 27th alone, over a dozen job postings were released in bulk. Liang Wenfeng has increased focus on the product, strengthening talent recruitment and exploration in the Agent direction, while also opening positions for "Search Algorithm Researchers." The responsibilities for this role include designing a new generation of general-purpose search engines for AGI and overseeing the large-scale implementation of LLMs in search scenarios. To突破 computing power bottlenecks, DeepSeek has begun recruiting "AI Supercomputing Cluster Operations Engineers." One of their duties is to "be responsible for the rapid delivery and deployment of new-generation computing resources, ensuring they are投入 production with high quality and performance." Experience in operating large-scale AI supercomputing clusters with over a thousand cards is a plus for the position. Also recruited in the same month were "Data Center Senior Delivery Manager" and "Data Center Senior Operations Engineer," with work locations in Ulanqab and monthly salaries up to 30,000 RMB. Liang Wenfeng had the foresight as early as 2021 to stockpile 10,000 NVIDIA A100 cards. "Liang is quite idealistic. Early on, with many cards on hand, he made them available to universities for research purposes," an AI practitioner revealed.

Huanfang's "Firefly One" Amid the complex chip landscape, Liang Wenfeng and DeepSeek carry significant expectations to "run domestic models on domestic computing power." Since DeepSeek's release, the architecture of domestic GPU chips has also begun to change. The V4 technical report mentions Huawei's Ascend in one section: the team validated the fine-grained Expert Parallelism (EP) solution on both NVIDIA GPUs and Huawei Ascend NPUs. "This indicates DeepSeek V4's inference path already possesses cross-computing-platform adaptation capabilities," an AI industry insider stated. Correspondingly, when introducing V4 pricing, DeepSeek included a line of greyed-out caption text with significant information: Limited by high-end computing power, the service throughput for Pro is currently very limited. It is expected that after the batch上市 of Ascend 950 super nodes in the second half of the year, the price of Pro will be大幅下调. These signals are encouraging. However, for domestic computing power to fully "hold up," time is still needed. "Domestic computing power is currently concentrated in the inference stage and limited to scenarios with high fault tolerance. There is still a distance from being used in the core pre-training stage, which consumes massive computing power." An AI industry practitioner offered a relatively conservative estimate: "There's at least an eight-to-ten-year gap to close."

Relentless Focus on Efficiency Returning to V4 itself, it is the most直观 outlet for observing Liang Wenfeng's model taste. Overall, the keyword remains unchanged: efficiency. DeepSeek-V4 pioneered a novel attention mechanism—compressing at the token dimension and combining it with DSA sparse attention. This achieves powerful long-context capabilities while significantly reducing computational and memory demands compared to traditional methods. Simply put, it brings a qualitative leap in long-text processing efficiency. Consequently, million-level context, once a "luxury" only affordable for closed-source flagship models, has suddenly become "dirt cheap." V4 has two versions: DeepSeek-V4-Pro (Expert Mode) and DeepSeek-V4-Flash (Fast Mode). The former is responsible for "strength," with官方 claiming performance comparable to top closed-source models; the latter is responsible for "economy," providing fast and cost-effective service. In terms of specialized capabilities, Liang Wenfeng's team focused on attacking the Agent direction. The V4 series has been adapted and optimized for mainstream Agent products like Claude Code and OpenClaw. Internally, it is claimed that in Agentic Coding evaluations, V4-Pro has reached the current best level among open-source models. "Based on evaluation feedback, the user experience is优于 Sonnet 4.5, with delivery quality接近 Opus 4.6 non-thinking mode, though there remains a certain gap with Opus 4.6 thinking mode." DeepSeek rarely disclosed the status of its internal use of Agentic Coding models. It should be noted that the version DeepSeek released this time is a "preview版"; the official version will have to wait a bit longer. "In the technical report, the team坦言 that V4's capability level still lags behind GPT-5.4 and Gemini-3.1-Pro, with a development trajectory roughly 3 to 6 months behind the leading closed-source models." Anchored to top-tier models, Liang Wenfeng remains执着求解: when computing power becomes increasingly expensive, can extreme architectural innovation continue to drive down computing power costs? "Not lured by praise, not frightened by slander, following the path of principle,端正 oneself with integrity." In the V4 release announcement, Liang and his team concluded with this statement,表明 their original初心.