Innovation that disrupts markets requires the courage to explore unique paths, not the exhaustion of copying rivals.
On July 31, DeepSeek revealed its latest model, the DeepSeek-V4-Flash official version, opening it for public testing. The architecture and parameter count match the previous version, but significant improvements in post-training have dramatically boosted its Agent capabilities. Combined with a price far below competitors, this release has sparked excitement among developers globally. An Australian AI professional posted on social media, praising DeepSeek's approach: they avoid heavy advertising and grand promises, instead disappearing for months to research silently before returning with a product that rivals top-tier offerings at a fraction of the cost.
As DeepSeek stands in the spotlight of global AI competition, its founder, Liang Wenfeng, has largely vanished from the public eye. Since early 2025, when DeepSeek gained fame with its R1 model, Liang has not granted media interviews and rarely appears in public. The company's website has long featured only a single public email address. However, 2026 marks a shift for this star company. After three years, it has opened its first external funding round and is aggressively hiring R&D staff, entering a new development phase. The goal of pursuing AGI remains unchanged, but the reality DeepSeek faces is different from a year ago.
No Interference, No KPIs
In July 2023, Chi Yu, a first-year doctoral student at a top Chinese university, joined DeepSeek as an intern researcher after three rounds of interviews. The wave of large language models ignited by ChatGPT was sweeping the globe, and just three months earlier, the quantitative fund High-Flyer had announced the creation of an independent AI research company. The month Chi Yu started, DeepSeek completed its formal business registration. With offices in both Hangzhou and Beijing, Liang Wenfeng frequently traveled between the two cities. Chi Yu worked there for a year, observing the company's early stages. His first impression of Liang was his humility. Chi Yu's expertise was in reinforcement learning, an area DeepSeek had not yet explored. "Liang told me, 'You teach me reinforcement learning,'" Chi Yu recalled.
Over the past two years, the outside world has often portrayed Liang Wenfeng as a tech idealist, but it is easy to overlook his skill as a manager who organizes research. Liang has explained his hiring logic: he values basic ability, creativity, curiosity, and passion over experience and resume. He notes that his team does not consist of "mysterious geniuses," but rather fresh graduates from top universities, senior PhD students, and young people who graduated a few years ago. Chi Yu fit this mold. In 2023, as the domestic large model race heated up with internet giants and startups vying for a spot, Chi Yu chose DeepSeek, which was not yet well-known. He told China News Weekly that he pursues AGI, admires heroism, and is confident in his abilities. DeepSeek attracted him because Liang suggested everyone could start from scratch to build a large model. He passed three rounds of interviews and an online test, which was very difficult. "Anyone who passed the interviews is highly capable," he said.
Liang Wenfeng trusts the people he hires, entrusting them with important tasks without interference or setting KPIs. As an intern researcher, Chi Yu independently led the early construction of DeepSeek's large model reinforcement learning infrastructure (LLM RL Infra). After completing the basic framework and gaining approval, Liang assigned more researchers to collaborate. During weekly meetings, the team reported progress on different modules of model training. Chi Yu recalled that Liang was sharp, focusing on technical details and persistently asking questions about what he deemed important. He felt that Liang was also interested in new things related to his familiar areas, but "this is also where he is humble; when he first encountered something, he would let us follow our own ideas." Chi Yu remembered that the team included many young researchers like him, either still in doctoral programs or recently graduated. Despite their limited experience, they had considerable autonomy. "It was really idealistic. With fewer people, the computing resources per person were abundant," he said. Liang has mentioned that such passionate young researchers crave scientific discovery more than money. However, that does not mean compensation was unimportant. In fact, DeepSeek's early salaries were competitive. Chi Yu did not disclose his exact salary but acknowledged it was higher than many peers. In February 2025, media noted DeepSeek's job postings, showing most positions offered monthly salaries above 20,000 yuan, core R&D roles could reach annual salaries of one million yuan, and AGI interns earned 500 to 1,000 yuan per day. This hiring philosophy was established during the High-Flyer Quant era. Liang has said that when High-Flyer was founded in 2015, the core team, including himself, had no quantitative experience. Key technical roles were also filled by fresh graduates and newcomers with one to two years of experience, and the main sales team started from scratch. He emphasized personal ability, stating, "It might not be a secret to success, but it's one of High-Flyer's cultures." A quantitative fund manager who knows Liang, as quoted by Reuters, revealed that a senior data scientist at High-Flyer could earn an annual salary of 1.5 million yuan, while similar roles at other companies were often below 800,000 yuan. Liang has mentioned that DeepSeek lacks a formal management system, motivating everyone through a shared goal. Chi Yu described the work atmosphere at DeepSeek as unrestricted hiring, flat hierarchy, mutual help, and freedom. During his year there, everyone was aligned toward AGI, and collaboration was seamless. "Wherever there was a shortage, someone would step in." He believes this atmosphere is linked to Liang's management style. "Liang is willing to listen to advice and will offer his own views in areas he knows well." After leaving DeepSeek, Chi Yu interned at large companies like Meta and noticed different organizational styles: Meta emphasizes individual heroism, but at DeepSeek, "everyone rides the rocket together, and lacking anyone is unacceptable."
Discipline as a Strategy
In 2002, Liang Wenfeng became the top scorer in the college entrance exam in Wuchuan City, Guangdong, with a score of 806. When sharing his study tips, he said, "Value self-study, don't just follow the teacher; have your own study plan, methods, goals, and standards." Looking back, many of Liang's choices seem to follow this approach. In the domestic and international AI industry, startups compete not only on models but also for attention. Founders often give interviews, remain active on social media, and pitch stories to capital markets. After the second half of 2024, the domestic "battle of a hundred models" entered its later phase, with industry discussions shifting to business models. Liang Wenfeng does not speak publicly, is not in a hurry to raise funds, and invests most resources in model development. Based on his few public statements and DeepSeek's development path, he does not follow industry mainstream or "consensus" but has clear judgments and goals. Since July this year, a transcript of Liang Wenfeng's internal exchange with investors has been widely circulated online. In it, Liang repeatedly uses the word "discipline," calling it a strategy. This is first reflected in the choice of an open-source approach. Over the past few years, Chinese and American large model companies have gradually formed two different paths. Companies like OpenAI have strengthened their closed-source model, while companies like DeepSeek insist on open source. Liang's judgment on open source is based on his understanding of industrial development. He believes that the value created by AI in the future will be enormous, "possibly accounting for 10% of human society's GDP," and no single company can monopolize such a market. Attempting to monopolize profits would ultimately lead to being eliminated by history.
The same logic applies to product pricing. Liang mentioned that DeepSeek does not pursue maximum commercial profit but maintains "reasonable" profit, "roughly enough to recover the cost of purchasing equipment in ten months." He frankly stated that giving up profits can make users feel goodwill and enhance the team's sense of identity. He shared that when the team initially set a higher price for the model, members were dissatisfied; after deciding to lower the price, "many people cheered" in the company group chat. The team feels that what they do is actually used by more people, not just for making money. In early 2025, a week after the release of DeepSeek-R1, daily active users exceeded 20 million, and at the end of the month, the peak surpassed 45 million. For a startup, this was the perfect window to capture traffic and quickly commercialize, but Liang did not focus on retaining users or rapid monetization. Since 2025, major companies like Tencent, ByteDance, and Alibaba have launched multiple rounds of competition around AI application entry points. However, Liang stated in the exchange that the team never thought of making DeepSeek the next super app and did not blindly invest in competing for individual users, "because there are bigger opportunities ahead, and the current ones might be insignificant." Chi Yu also felt this. He recalled that the team has always focused on AGI, "so we are not so concerned about product matters. Liang also concentrates resources on the most important things, like everyone prioritizing pre-training." This judgment was also reflected in earlier arrangements. Before the wave of large models truly arrived, High-Flyer Quant had already started investing in AI infrastructure. Around 2020, High-Flyer invested nearly 200 million yuan to build the deep learning training platform "Firefly One"; two years later, it spent about 1 billion yuan on "Firefly Two," equipped with about 10,000 Nvidia A100 GPUs. According to reports, at that time, few domestic companies besides internet giants had such a scale of AI computing power reserves. Liang later explained that there was no complex commercial logic behind this, but rather a curiosity about the boundaries of AI capabilities. For outsiders, ChatGPT is a landmark moment, but for more AI practitioners, the 2012 victory of AlexNet in the ImageNet competition had already demonstrated the feasibility of "big data plus big computing power creating miracles." Since then, Liang had been consciously deploying computing power.
Of course, Liang's ability to maintain his own pace is also supported by practical conditions. High-Flyer Quant has long had stable cash flow, providing DeepSeek with R&D budgets. Liang has stated that High-Flyer, as one of DeepSeek's funders, has sufficient R&D budgets, with annual budgets of hundreds of millions for public welfare, which can also be redirected to R&D if necessary. But money doesn't explain everything. An anonymous large model practitioner told China News Weekly that since 2023, after OpenAI's success, many domestic entrepreneurs have expressed a desire to become the "Chinese version of OpenAI." After Anthropic gained attention, companies mimicking its path appeared, and after DeepSeek became popular, similar followers emerged. In his view, this reflects a common phenomenon in Chinese AI entrepreneurship: absolute obedience to the strong and a tendency to copy already proven successful paths. Liu Zhiyuan, a professor in the Department of Computer Science at Tsinghua University and an AI entrepreneur, believes that Liang Wenfeng's most noteworthy aspect is not the specific technical route he chose, nor is every judgment of his correct, but his ability to maintain independent judgment, even persisting in what he thinks is right when others are skeptical. What disruptive innovation needs is the courage for differentiated exploration, not the "involution" of copying.
The Cost of Discipline
In June this year, DeepSeek completed its first external financing round since its founding, raising over 50 billion yuan (approximately $7.4 billion). This is the largest single-round financing in the history of China's AI large model industry, marking the official opening of the company to external capital. Why did Liang Wenfeng change? In the view of many industry insiders, the core reason is not a lack of money. The aforementioned large model entrepreneur told China News Weekly, "If it's just for training models, DeepSeek probably doesn't have such a large financing need." He analyzed that, given DeepSeek's current R&D model, High-Flyer's stable long-term income is sufficient to support model training, and DeepSeek itself is known for low-cost training, without the massive investment in C-end products like many peers. Outside speculation suggests the real pressure comes from talent. For a company that has not raised funds or gone public for a long time, the options held by employees are difficult to form clear value. When the entire industry starts a new round of AI talent competition, it is hard to retain top researchers for the long term with idealism alone. Over the past year, DeepSeek has become one of the most important sources of talent for the domestic AI industry. After R1 became popular, major companies joined the talent war. Xiaomi reportedly poached Luo Fuli with an annual salary of 10 million yuan; Guo Daya, a core contributor to models like DeepSeek V3 and R1, joined ByteDance's Seed team; Wang Bingxuan, the core author of the first-generation large language model, joined Tencent's Hunyuan; Ruan Chong, head of multimodality, left to join YuanRongQixing; and another core member, Wei Haoran, has also resigned. Meanwhile, companies like ByteDance, Tencent, and Alibaba have all issued recruitment slogans like "no cap on salary" and "unlimited treatment," intensifying the AI talent competition. Liang Wenfeng also admitted in the internal exchange that the biggest risk for DeepSeek is team stability. He stated bluntly, "As long as I can maintain team stability, I can definitely achieve AGI." After this financing, this risk has been somewhat alleviated. "The options everyone got are still quite a lot, and the amounts are relatively large." In fact, Liang Wenfeng is not completely opposed to capital, but worries that capital will change the company's development pace. He mentioned that the early team also communicated with different funders, but many VCs had concerns about doing research, as they had exit needs and hoped to commercialize products quickly, which conflicted with the company's priority of doing research first. It is worth noting that even when deciding to open up financing, Liang still tries to keep the initiative in his own hands. According to media reports, in this round of financing, Liang Wenfeng personally invested about 20 billion yuan, accounting for about 40% of the total financing, making him the largest single investor. Although investors like Tencent, CATL, JD.com, and NetEase participated, they do not have voting rights, and their shares have a five-year lock-up period. However, with the completion of financing, DeepSeek inevitably faces new changes. In late June, DeepSeek launched its largest-ever recruitment drive, explicitly stating that it will at least double the size of each department, with positions covering algorithms, R&D, products, operations, data engineering, and functions. Unlike the past when the research team was the core, this recruitment also includes product-related positions. The aforementioned large model practitioner told China News Weekly that in the past, Liang Wenfeng could focus almost entirely on research, but with the completion of financing, the company needs to face clearer commercial expectations. "Before, there was no rush to consider products, but in the future, it will be necessary to strengthen product, service, and commercialization capabilities. This is inevitable as the company reaches this stage." A headhunter working for DeepSeek told China News Weekly that DeepSeek does not lack applicants, and recent hiring standards remain very high, preferring quality over quantity. However, according to the laws of enterprise development, the culture formed during the small team period often faces challenges as the scale expands. Over the past three years, Liang Wenfeng has shaped DeepSeek's unique working style through personal judgment and team culture. But as the company enters a new phase, how to extend the tacit understanding formed by idealism and high trust to a larger organization may be the real challenge Liang Wenfeng will face. (Chi Yu is a pseudonym to protect the interviewee's privacy.)
Comments