As early as June 25, as soon as the first round of 51 billion yuan in financing was settled, DeepSeek quickly issued a recruitment announcement, clearly proposing to “at least double the size of all departments” and opening 33 positions at one time, covering seven major sectors including technology research and development, computing infrastructure, product implementation, cutting-edge algorithm research, and comprehensive functions. The work locations are in Beijing and Hangzhou, and it even recruits data center infrastructure positions in Ulanqab – where it builds its own computing cluster.

The pace is faster in September. Cui Tianyi, head of the DeepSeek Harness team, released about 150 recruitment spots on social platforms, saying the demand was “unprecedented.” The destinations of these 150 places are very concentrated: almost none of them are for AI research positions. They are all for server-side development engineers and agent elastic computing R&D engineers. They are clearly targeted at senior back-end personnel with 2 to 10 years of experience. Cui Tianyi’s explanation is very straightforward. After the team grows, what is needed is to turn the model capabilities into a stable and callable engineering system. At the same time, three positions of researchers, R&D engineers, and product managers in the Agent Harness direction are open for a long time. Both internships and full-time positions are available, and the jobs are in Hangzhou and Beijing.

A company with only 300 to 400 people doubled its headcount, but prioritized backend and engineering rather than algorithms. The direction was clear: it wanted to shift from “making good models” to “stablely handing models over to more people for use.” Cui Tianyi also specifically clarified the rumors of “not recruiting foreigners” on social platforms – just like similar companies in the United States require employees to work in English, DeepSeek wants people who can collaborate, regardless of nationality.

V4系列

DeepSeek 2026 will be updated almost quarterly. On April 24, the V4 preview version was launched and open sourced simultaneously, focusing on million-word context: V4-Pro is a 1.6T total parameter, 49B activated MoE, and its performance benchmarks against top closed-source models; V4-Flash is a 284B total parameter. On August 13, the official version of V4 Pro was fully launched on the APP, web page and API, and users can call it through “expert mode”. On September 10, the smaller and faster V4.1 Flash was released. As the smallest model in the new architecture family, it natively supports multi-modal visual understanding and adopts an asymmetric architecture with 552B parameters.

Ecologically, DeepSeek has previously announced that it will open source its internal inference framework DeepInfer and be compatible with two computing power platforms, Nvidia and Huawei Ascend. This step is generally regarded as its major investment in the adaptation of domestic computing power. OpenRouter’s data gives the most direct market reaction: DeepSeek’s model has ranked first in the platform’s weekly call volume list for three consecutive weeks. On October 23, it was also reported that DeepSeek and MiniMax had completed registration and completed compliance procedures for open source models and overseas authorized use.

In August, it also specifically recruited high-paying talents for game development and 3D rendering. The positions include Code Agent data engineer (game direction), Web 3D and rendering talents, and candidates are required to “do not have to be proficient in model training, but must have actually done games.” This is consistent with the participation of Perfect World and Century Huatong in this round of financing.

Stay in a hotel in Hangzhou and rent an apartment in Beijing

DeepSeek’s main company is registered in Gongshu District, Hangzhou, but its real research and development center is in the Raycom Information Center in Haidian District, Beijing. DeepSeek’s Beijing subsidiary has rented out an entire floor of it, covering an area of ​​about 2,100 square meters. According to multiple people close to the company, most DeepSeek R&D personnel are actually in Beijing.

Hence Liang Wenfeng’s unique twin-city life. As the founder of a Hangzhou company, he lived in hotels in Hangzhou for a long time. In Beijing, where R&D personnel gather, he always rented a house, and the room layout was always relatively simple. In the impression of people around him, this man who achieved financial freedom by relying on quantitative trading under the age of 30 would wear the same clothes for many days. He was lean and kept active. His well-known hobbies were outdoor activities such as hiking.

This simplicity is completely different from his past work. Liang Wenfeng founded Huanquan Quantitative with his classmates from Zhejiang University in his early years. In 2019, he spent nearly 200 million yuan to build a self-built GPU cluster (1,100 cards). In 2021, he spent an additional 1 billion yuan to stock up on about 10,000 A100s. He himself admitted that he simply could not use so many cards for quantification. It was this batch of computing power saved in advance that later became DeepSeek’s financial resources to continue training models under chip control. Now the company is valued at 500 billion. He is still renting a house and basically invests his money in computing power and people.

People who know him well say that Liang Wenfeng dares to make big bets on the things he sees, but his life is very simple. Huanfang Quantitative has been ranked at the forefront of domestic quantitative private equity all year round. This business is the capital that he dared to spend about 20 billion yuan in the first round of financing and can live in a rental house for a long time. Here, money is not used for consumption, but is replaced by more cards, cheaper computing power, and more advanced bets. Compared with those peers who rely on rounds of financing to survive, DeepSeek has not been on a runway from the beginning.

Beijing and Hangzhou

Putting the office space, recruitment positions and investment layout together, DeepSeek’s division of labor in Beijing and Hangzhou can be seen relatively clearly.

Hangzhou is where the company’s main body and headquarters functions are located, hosting industrial and commercial, capital operations and some engineering teams. It is also its “front desk” for external financing and connecting listed companies and industrial capital. Beijing, with its entire office area of ​​​​the Raycom Information Center, has gathered the most intensive algorithms and core research and development forces, and is the main battlefield for model training, cutting-edge exploration and agent systems. Liang Wenfeng himself has been based in Beijing for a long time, which also means that the company’s technical brain and decision-making focus have actually moved north.

Ulanqab is outside this territory and is responsible for the backend of its self-built data center and computing infrastructure. This is why there will be data center infrastructure positions in the round of recruitment expansion in June. Financing has expanded from major Internet companies to semiconductors (SMIC Juyuan), automobiles (Geely), and energy hardware. This corresponds to this asset-heavy path of “computing power-chip-terminal”.

物理世界

In recent months, Liang Wenfeng has communicated intensively with many embodied intelligence entrepreneurs to discuss the direction of interaction between robots, large models and the real world. On the investment side, DeepSeek has already made three investments, spending about 141 million yuan to invest in Yushu Technology, which is involved in robot body and motion control, and plans to cooperate in the development of robot AI models; it plans to lead the investment in the Winverse AI seed round (tens of millions of dollars, post-investment valuation of about 400 million U.S. dollars) founded by former Tencent AI scientist Hu Han; it also participated in the financing of Yu Tao, a former member of Tencent’s Juewu team and the person in charge of the humanoid basic model of Zhuji Dynamics.

From running models in the cloud to making robots move in the real world, embodied intelligence requires continuous consumption of computing power, building test environments, and binding hardware entities—this is probably why the financing amount has increased from 50 billion to 100 billion. The entry of semiconductor capital such as SMIC into the market also makes people speculate whether it will go further in adapting the bottom layer of domestic chips and even develop its own chips.

Related Reading

• Audi moved a “rescue” from Chery2026-10-28
• The market value of the six luxury goods giants has shrunk by nearly 2 trillion since the beginning of the year, with LVMH losing the most blood.2026-10-27
• Original “Nothing to film” Jia Nailiang sold 570 million yuan in a single day, but the company behind him suffered a huge loss of 3.7 billion yuan in five years2026-10-27
• Original bit system is eating up our time crazily2026-10-27
• EDonkey, can you cross the river by touching a motorcycle?2026-10-26