Summary:
Over 30 institutional investors, with more than 5 billion yuan in capital, have poured into a company that hasn't yet achieved mass production or significant sales revenue in just two and a half years.
Produced by Phoenix Network Technology's "Fengyan Observation"
On August 12th, at a SF Express warehouse in Shenzhen, Zibianliang Robotics held a fully livestreamed test of its logistics sorting capabilities. The officially announced results were impressive: 1,816 random parcels sorted in one hour, with a success rate exceeding 98%, averaging 1.98 seconds per item. Some reports pointed out that this number was about 45% higher than the 1,248 items/hour industry benchmark previously set by the embodied AI company Figure AI, with costs allegedly cut by 70%.
Controversy arrived even faster than applause. The Douyin blogger "Robotics Detective" posted a dissection video: Figure AI's score came from a 200-hour non-stop livestream, with multiple robots taking turns, accumulating a total of about 249,000 items sorted, averaging 1,248 items/hour, while Zibianliang only broadcasted for 1 hour. Furthermore, Figure used full-size humanoid robots, while Zibianliang adopted a simpler solution of dual robotic arms + standard grippers.
"It's like Wang Qian riding an electric bike and racing a marathon world champion in a 100-meter sprint, then winning and mocking the other's running shoes for being too expensive," said the blogger "Robotics Detective" in the video's caption.
This livestream perfectly encapsulates the concentrated operations of Zibianliang over the past half year—using the most attention-grabbing actions to repeatedly tell the most grandiose story, speeding down the fast track of capital-driven hype.
From Quantitative Fund to Embodied "OpenAI": 5 Billion Yuan in Financing in Two and a Half Years
Zibianliang's story begins with a young man named Wang Qian.
Wang Qian, born in 1988, holds bachelor's and master's degrees from Tsinghua University and a Ph.D. from the University of Southern California, researching robot learning and human-computer interaction. He has a frequently mentioned label: one of the earliest scholars globally to introduce attention mechanisms into neural networks, with results published in international conferences concurrent with Google. This is also why Zibianliang has been able to consistently attract attention from giants: the attention mechanism is a core module of Transformer, which is the foundation of all large models today.
After his Ph.D., Wang Qian neither stayed in academia nor directly entered a major tech company. He founded a quantitative fund in the US, using AI for trading. At the end of 2023, he dissolved the fund and returned to China, founding Zibianliang Robotics in Shenzhen.
Wang Qian's entrepreneurial partner and Zibianliang's co-founder and CTO, Wang Hao, holds a Ph.D. in Computational Physics from Peking University and formerly served as the algorithm team lead for the Fengshenbang large model team at the IDEA Research Institute, leading the development of China's first multimodal open-source large model "Taiyi." One understands underlying architecture, the other understands large-scale training engineering—this combination was very persuasive in Zibianliang's early startup phase.
Zibianliang's technical path differs from traditional robotics. Instead of piecing together perception, planning, and control modules, it has from day one pursued an end-to-end unified pre-training model, putting vision, language, action, and physics prediction into a single network for joint training, allowing the model to understand the entire physical world. In April 2026, the company released its embodied large model WALL-B based on the "World Unified Model" architecture, claiming to break the industry convention of pre-training plus post-training.
Zibianliang's products are also quite different from other companies. The robot product showcased in the livestream at the SF warehouse did not take a humanoid form, nor was it a mobile dual-arm body + dexterous hand. Instead, it utilized a fixed workstation + gripper. Zibianliang's explanation was that such an architecture is closer to the behavioral habits of human hands and can achieve faster speeds.
The advantage of the technical path Zibianliang chose is a high ceiling; the downside is that it burns money fast. Full-stack self-research means models, data, bodies, dexterous hands, and data collection systems all have to be built from scratch, leading to long R&D cycles and huge investments. Fortunately, currently, capital's patience is just sufficient. Zibianliang has achieved a rare financing speed in Chinese venture capital history.
In May 2025, Meituan Strategic Investment led the Series A round, with several hundred million yuan. Four months later, Alibaba Cloud and Guoke Investment followed with the Series A+ round, nearly 1 billion yuan. In January 2026, ByteDance and Sequoia China jointly led the Series A++ round, 1 billion yuan. On April 20th, Xiaomi Strategic Investment and Sequoia China led the Series B round, nearly 2 billion yuan. From May to June, Series B+, B++, and C rounds were densely closed within just over two months, with valuations jumping directly from 10 billion to 20 billion yuan.
Meituan, Alibaba, ByteDance, and Xiaomi—four internet giants—appeared as lead investors in different rounds. Zibianliang became the only embodied AI enterprise in China to simultaneously secure backing from these four major firms. Sequoia China participated from the A+ round onward, investing in every subsequent round. Top-tier institutions like IDG, Source Code Capital, Dachen, and CICC all participated. State-backed funds such as China Mobile Chain Leader Fund, National Artificial Intelligence Industry Investment Fund, CDB Financial, China Insurance Investment, and Shenzhen Investment Holdings also joined successively. Industrial capital from Honor, Chery, 58 Group, Shenyang Automobile also appeared on the shareholder list.
Over 30 institutions, with more than 5 billion yuan in funds, poured into a company that hasn't yet achieved mass production or significant sales revenue in just two and a half years.
Frenzied 5 Billion Yuan Financing: Where Is Zibianliang's Money Burning?
A company that hasn't yet achieved mass production or significant sales—what qualifies it to secure 5 billion yuan in financing, and where will this money be burned?
First, look at the valuation model. In the same embodied intelligence track, Unitree sells mature products, selling over 5,500 humanoid robots in the whole of 2025, with revenue nearing 1.7 billion yuan. These robots can dance and even box; consumers and investors have a clear sense of Unitree's product capabilities. In contrast, Zibianliang's selling point is its model capability and the imagination space for future commercial value. The market is expecting the actual deployment results of the WALL-B model and the scaling capabilities of its data factory.
Model capability, being so crucial, naturally becomes a major area where Zibianliang burns money. Full-stack self-research of an end-to-end large model requires continuous investment in computing power and talent costs. Zibianliang's team comes from Tsinghua, Peking University, USC, and top hardware companies, commanding high salary levels. After releasing WALL-B in April 2026, Zibianliang successively open-sourced the WALL-OSS-0.5 base model, launched the XRZero-G0 embodiment-free data collection solution, and the X-Tokenizer cross-modal action tokenizer. Its product lines are widely spread, each requiring continuous investment.
Data collection supporting model R&D is the second area of heavy spending. The biggest difference between embodied large models and language large models is that language models can crawl massive amounts of text from the internet, while robots must collect data in the real physical world. Real-world data is expensive and slow to collect. Zibianliang built China's first large-scale robot data factory in Shenzhen and launched an embodiment-free data collection solution, claiming to reduce data collection costs by about 95%.
Zibianliang's home data collection is another point of considerable controversy. Recent media reports indicate that Zibianliang is currently conducting a "Home Data Collection Family Recruitment," requiring houses over 90 square meters, with three-bedroom layouts preferred, stable home network, and a "real, rich home atmosphere." From 9 AM to 7 PM daily, operators bring a complete robot unit and a set of accompanying consoles into the home for operation, which residents can supervise throughout. The maximum subsidy is 3,000 yuan per month, for up to two consecutive months, capped at 6,000 yuan per household.
Simply put, Zibianliang is paying users to allow robots into their homes to collect data.
This forms a stark contrast with Zibianliang's previous commercialization attempts. In March 2026, Zibianliang partnered with 58 Daojia to launch a robot cleaning service, priced at 149 yuan for 3 hours, available in Shenzhen and Beijing. The form was human-robot collaboration: a cleaning aunt leads a team, with the robot acting as an intern, doing tasks like folding clothes, tidying toys, wiping tables, scooping cat litter, and organizing shoes. Leveraging the "internship" opportunity, it enters real-world home scenarios to collect data.
According to media reports, Zibianliang's smart cleaning service on the 58.com platform offered 8 slots per day, with appointments booked half a month in advance. It looked popular, but the actual experience wasn't reliable. A blogger who spent 149 yuan found that the robot took ten minutes to fold one piece of clothing, only tidying the shoe cabinet, dining table, and sofa in three hours, with slow speed and loud noise. Gan Ruyi, a partner at Zibianliang, admitted in an interview with National Business Daily that the speed and noise indicated current technology still had room for improvement, and home scenarios were far more complex than anticipated.
However, by the second half of 2026, this 149-yuan robot cleaning service had faded from the company's promotional rhetoric.
From asking users to pay 149 yuan to invite a robot into their home, to the company subsidizing 3,000 yuan per month for users to allow a robot in, Zibianliang has verified with real money that current robots entering homes don't offer much service value and are more of a disturbance to users. The model packaged as a paid service didn't work; it's better to directly account for this expense as data collection cost.
Progress in industrial scenarios is also limited. The company disclosed that its model has entered the production lines of Brilliance BMW seat and interior suppliers and collaborated with SF Express on logistics sorting. These pilot projects still have a noticeable distance from mass production and delivery.
Behind the Secret Submission: The Shackles of High Valuation
On August 5th, the South China Morning Post, citing two informed sources, reported that Zibianliang had secretly submitted an IPO application to the Hong Kong Stock Exchange. Additionally, its third-generation robot is planned for release in the fourth quarter of 2026. This is just two and a half years after the company's founding.
The timing is intriguing. Not long ago, Unitree Technology just completed its A-share subscription at an issue price of 150.8 yuan, with its total market value once soaring to 400 billion yuan, enjoying the title of the "first humanoid robot stock" on the A-share market. Peers are starting to speak with shipment volumes and revenue figures; Zibianliang must prove itself in new ways.
High valuation is a double-edged sword. Over 30 institutions used real money to push Zibianliang to 20 billion yuan. They tolerate the initial crazy burn of money to build data factories and conduct pre-training, but their patience is limited. As the hype of financing news fades, investors will begin scrutinizing Zibianliang's financial statements.
Next, there is a series of very specific questions that Zibianliang needs to address one by one: Through what revenue form will model capabilities ultimately be realized? Is it body sales, solutions, software services, or data licensing? For each new customer added, how many engineers are needed for on-site debugging? If entering each new scenario requires re-collecting data and fine-tuning the model, then it's not a platform company but a project-based outsourcing firm. Project-based companies don't have marginal cost reduction effects; no matter how hard they work, they can't support a 20 billion yuan valuation.
The competitive landscape in the embodied intelligence industry is also constantly changing. The local rival Zhi Ping Fang also has a 20 billion yuan valuation, has secured nearly 500 million yuan in orders, and is called by the industry "the company most like Tesla" in the embodied AI field. Unitree proved with 1.7 billion yuan in revenue and profitability that robot bodies can be a business. Zhiyuan has rolled off its 15,000th general-purpose embodied robot. Leju is applying for a listing on the ChiNext board with 258 million yuan in revenue. While peers are successively submitting answer sheets with revenue and shipment volumes, Zibianliang's score is still unknown.
The risks of the chosen technical path are equally real. End-to-end unified pre-training models require large investments and long cycles. Views within the industry point out that many so-called self-developed models are actually repackaged open-source models after post-training. Zibianliang repeatedly emphasizes starting pre-training completely from scratch—this is a technical differentiation and a market stance distinction. Full-stack self-research itself does not automatically become a barrier; it only generates economic value when it can reduce training data for new tasks, compress deployment cycles, and lower delivery costs.
A practitioner in the embodied track told Phoenix Network Technology, "I don't believe the current domestic environment can support building a company completely benchmarked against Figure. Past experience suggests it's still more suitable to first achieve deep penetration in a single point." This implies that Zibianliang, aiming for everything, currently faces significantly greater pressure.
According to feedback from multiple former Zibianliang employees posting on social media, the company's internal work pressure is immense. Someone posted that due to high work intensity and excessive overtime, they experienced health issues, but even when taking sick leave, they would be questioned by superiors about "whether they were really sick." This individual believes that joining Zibianliang had a certain impact on their career development. According to their statement to Phoenix Network Technology, they joined Zibianliang two months ago and have already left.
To some extent, that August sorting livestream was a concentrated exposure of Zibianliang's anxiety. It needed a sufficiently viral number to support the narrative of its 20 billion yuan valuation, so it chose to benchmark a 1-hour sprint against a 200-hour marathon, fixed robotic arms against humanoid robots. The "45% surpassing Figure AI" result looked eye-catching, but upon dissection, were there elements of careful design? Zibianliang still needs to answer with stronger capabilities.
Figure AI's valuation is 39 billion USD, approximately 280 billion RMB. Zibianliang's 20 billion RMB valuation is about one-fourteenth of Figure's. This act of benchmarking itself indicates that Zibianliang needs a sufficiently large reference point to anchor its imagination space.
In two and a half years, Wang Qian has brought Zibianliang from an office in Shenzhen to the card table with a 20 billion yuan valuation. The backing of four major internet giants, the real money from over 30 institutions, the technical path of an end-to-end large model—these are all real.
On the flip side of these facts, one must also see that embodied intelligence is a long and arduous track. Model capabilities ultimately need to translate into productivity that customers are willing to pay for continuously. Zibianliang has already assembled models, data, bodies, and capital on the same map. The third-generation robot and future public operational data will test whether these links can form a closed loop.
The 20 billion yuan valuation is buying the future. When that future needs to be cashed out with revenue, Zibianliang's answer sheet has only just begun to be written.
This article is from WeChat Official Account: Phoenix Network Technology , Author: Phoenix Network Technology, Editor: Dong Yuqing
