The capital blitz has begun, with Lei Jun and Jack Ma jointly leading the investment in a rare co-appearance.
Qianxun Intelligence has once again fully accelerated its fundraising pace.
On April 7, 2026, Qianxun Intelligence announced the completion of a new round of financing amounting to RMB 1 billion. This round was co-led by Shunwei Capital and Yunfeng Capital, with significant participation from Dacheng Capital, a leading RMB-denominated fund, Galaxy Yuanhui, Turing Fund, Xinding Capital, and Gengxin Capital.
This is already its second major funding round within 30 days. Just recently in February, the company completed a funding round of nearly RMB 2 billion. Combined, the two rounds bring the total funding to RMB 3 billion.
More interestingly, this round saw a highly talked-about pairing: Lei Jun (Shunwei) and Jack Ma (Yunfeng) co-leading the investment for the first time in the embodied AI space.
In the past, they each correctly bet on key cycles such as mobile internet, e-commerce, smart hardware, and cloud computing. This time, by joining forces to invest in robotics—particularly in the still-early field of embodied AI—they signal that this direction is moving from technological imagination to capital consensus, entering a ranking and elimination phase backed by industry giants and fueled by concentrated capital.
Qianxun Intelligence was founded in January 2024 by continuous robotics entrepreneur Han Fengtao, leading AI scientist Gao Yang, and robotics globalization pioneer Zheng Lingyin.
Founder and CEO Han Fengtao previously served as co-founder and CTO of Rock Robotics, where he led the delivery of nearly a hundred robot models, demonstrating deep expertise in engineering and mass production. Co-founder Gao Yang graduated from the University of California, Berkeley, where he studied under computer vision pioneer Trevor Darrell and is now an Assistant Professor at the Institute for Interdisciplinary Information Sciences at Tsinghua University. Under his leadership, the team open-sourced the Spirit v1.5 model, which surpassed the leading U.S. model Pi0.5 on the RoboChallenge leaderboard, becoming the first Chinese open-source embodied AI model to reach the top. Co-founder Zheng Lingyin is a pioneer in the international expansion of industrial robots, having built the overseas division from scratch and led her team to deeply penetrate multiple international markets, rapidly achieving commercialization outcomes.
The three founders collectively bring core competencies in AI, robotics, and commercialization, forming a rare “hexagonal warrior” team—an arrangement that underpins the company’s ability to secure 3 billion RMB in funding within 30 days and attract heavy investments from both Sequoia Capital China and Yunfeng Capital. This unique combination has endowed Qianxun Intelligence since its inception with both world-class technological foresight and a strong foundation for commercial execution.
Han Fengtao previously pointed out that in 2026, the competition will come down to data scale and model performance. This year, the most important goal is not expanding use cases, but making embodied models among the top three globally. To achieve this, sufficient funds must be available.
Therefore, this blitz-style continuous funding essentially trades capital density for a time advantage, rapidly scaling resources to create a performance gap and secure a top-tier position early. Meanwhile, existing shareholders continuing to invest in this round signals that investors have shifted from observation and validation to accelerated commitment.
So, what exactly has enabled Qianxun Intelligence to secure this accelerated entry ticket? How deep has its moat become?
The underlying logic of capital investment has been validated: a path more akin to large models has succeeded.
Why are investors willing to continue increasing their bets? The answer: the model has already provided a阶段性answer.
In January this year, Qianxun Intelligence open-sourced the embodied model Spirit v1.5, which outperformed the then strongest open-source model, Pi0.5, in public evaluations.
But what truly resonates with capital is the inflection point in the capability curve.
Spirit v1.5 has demonstrated relatively stable zero-shot generalization capabilities—performing complex tasks such as wiping, opening and closing hinges, and manipulating flexible objects without additional training.
In other words, robots are no longer just learning individual tasks—they are gaining the ability to transfer skills across tasks, offering a glimpse into the potential of embodied intelligence to liberate human productivity.
Behind this lies a technical approach highly similar to that of large language models (LLMs): scale up the model, feed it ample data, iterate continuously, and trust in the emergence of capabilities.
Specifically, Spirit v1.5 is an end-to-end Vision-Language-Action (VLA) unified model. It does not focus on reconstructing every detail of the world or emphasizing an explicit intermediate world simulation; instead, it directly learns the mapping from perception to action.
The training approach is also highly LLM-inspired. Instead of text data, robotic data is used. First, pre-train on massive amounts of internet videos to build a foundational understanding of the world, then align with real-world interaction data—first acquiring generalization capabilities, then refining for specific tasks.
As a result, it achieved stronger generalization performance with lower computational power and parameter scale.
Just a few days ago, this path also received a "resonant response" from peers in Silicon Valley.
On April 3, Silicon Valley embodied intelligence company Generallist AI released its foundational model GEN-1, validating the Scaling Law in embodied intelligence using 500,000 hours of real-world physical interaction data. How powerful are the results?
These robots have significantly increased the average success rate of multiple physical tasks from 64% to 99%, performed at nearly human speed—about three times faster than existing state-of-the-art systems—and can improvise on the fly. Even more remarkably, acquiring each new capability requires only about one hour of robot data.
Company CEO Pete Florence noted that what’s happening in robotics right now is similar to when people first opened GPT-3 and asked it to write a brand-new limerick.
Similar observations have also been validated by the Qianxun team. “Our team has also discovered the Scaling Law in embodied intelligence—every time the data increases tenfold, we gain another nine in the results,” Gao Yang once described the steepness of this curve. We are currently at the Scaling Law moment for embodied intelligence; because robotic data is harder to obtain, I believe the robot equivalent of GPT-4 will take longer—perhaps 4 to 5 years.
Capital is being poured into a technology pathway that has already been preliminarily validated, offering higher cost-efficiency and greater scalability potential.
Data engine, the key to path validation
In the field of embodied intelligence, there is nearly universal consensus that data collection is a fundamental bottleneck.
Large models can ingest the vast corpus of internet data, but robots cannot—in the world of physical labor, there is no Wikipedia. On the surface, everyone is competing over models, but the deeper competition is really over data engines. “To achieve scaling, we will spare no effort,” Pete Florence frankly stated.
Given belief in the Scaling Law, what kind of data system can be acquired at low cost, scaled continuously, and possess sufficient diversity?
Previously, robotic general models with success rates exceeding 90% relied on extremely expensive and hard-to-scale datasets from remote operation (e.g., Physical Intelligence). However, Generallist AI has自主研发 developed "Data Hands"—a two-finger wearable device worn on the wrist—that transforms human hands into robotic grippers to collect visual and sensory data.
As a result, the progress of GEN-0 and GEN-1 demonstrates that this data engine can also achieve high-level proficiency—it did not use robotic data, but instead relied solely on data generated by millions of activities performed by humans wearing low-cost wearable devices.
Qianxun Intelligence is also advancing a scaling roadmap centered on diversity.
In terms of hardware, Qianxun also adopted a wearable solution—but took it further. To enable the model to learn human-level fine manipulation, they implemented a three-finger design—the intelligent system features 26 degrees of freedom, with force sensors integrated into each joint and equipped with a dexterous three-fingered hand. However, the technical challenges have significantly increased: the three-finger structure presents higher demands for degrees of freedom, more precise force control, and more complex motion mapping in wearable data acquisition.
Currently, Qianxun's wearable devices have been upgraded to the fifth generation, with data availability increasing from 30% to 95%, while costs have been reduced to approximately one-tenth of those for teleoperation.
Note that, unlike Generallist AI, which relies entirely on wearable data, Qianxun builds a multi-source integrated data engine.
During the pre-training phase, Qianxun Intelligence actively integrates internet videos alongside vast amounts of wearable data to acquire general knowledge and foundational capabilities. Subsequently, real-world teleoperation data is introduced to perform fine-grained SFT (supervised fine-tuning), enhancing the model’s performance on actual tasks. Finally, reinforcement learning is applied for further optimization: the model continuously roll-outs in real-world environments, generating new data that feeds back into its training.
To date, Qianxun has collected over 200,000 hours of real-world interaction data from multiple sources, including internet video, teleoperation, and wearable capture, with this number continuing to grow rapidly and projected to exceed 1 million hours by 2026. By April 2026, Qianxun’s data collection team will also reach a scale of 1,000 members.
It is worth noting that Qianxun's understanding of data also underwent a fundamental transformation.
They moved away from the industry’s traditional, highly polished scripted data approaches and adopted a more open, diverse collection paradigm: rather than rigidly defining action paths, they focus on task objectives and allow the process to unfold naturally—embracing failure, allowing disruptions, and continuing until completion.
The resulting change is fundamental: the model no longer learns how to perform this specific task, but rather how to handle similar situations. Under the same data scale, this data distribution significantly improves the model’s transfer efficiency while reducing its dependence on computational power.
"Laying eggs along the way"—real-world scenario data feeds back into the model.
In Qianxun’s data engine, what truly determines whether the flywheel can spin is not just the data sources, but the ability to continuously roll out in real-world environments.
Han Fengtao once summarized that moving toward real-world scenarios is about acquiring the fuel for model evolution—data. Commercialization makes this data acquisition process sustainable and scalable.
Behind this lies a clear divergence in the U.S.-China pathways. In the U.S., some companies can sustain long-term investment in foundational models, trading time for higher capability ceilings; in China, without a demo or tangible deployment signals, it is difficult to secure continued funding. Most companies that survive—and even thrive—tend to adopt a more balanced approach.
The path to general AI is a long, snowy slope—it’s impossible to wait for models to mature before seeking applications. Only by deploying robots into real production environments and involving them in actual business operations can we leverage the massive amounts of data generated to feed back into and continuously evolve the models.
As China’s first embodied AI company to transition diverse data collection approaches from theory to engineering and scaling, and to achieve dual validation in real commercial scenarios, Qianxun Robotics adheres to the “eggs along the way” philosophy. They begin with controllable environments, prioritizing entry into industrial and service sectors—domains characterized by relatively stable structures, clear task boundaries, high profitability, and willingness to pay—thereby validating model capabilities while supporting company operations.
For example, in retail scenarios, Qianxun’s collaboration with JD.com (also an investor) is deepening. “Xiao Mo” has been deployed at JD Mall as a barista. While performing service tasks, the robot simultaneously collects multimodal sensory data, joint motion trajectories, and precise force feedback information.
These "expert-level data" from real retail environments will be directly used to train and fine-tune embodied models, creating a positive feedback loop of "data collection—model iteration—capability enhancement."
Qianxun Intelligent Robot has officially begun work at JD Mall as a barista.
Both parties also plan to extend embodied intelligence to additional retail sub-sectors, including digital appliance guidance, automated inspection and cleaning, and more. Meanwhile, JD Pharmacy is viewed as a core breakthrough area, where robots will participate in high-precision tasks such as automated sorting and accurate dispensing, exploring unmanned smart pharmacy solutions.
Before entering the JD Mall, Qianxun had already undergone validation in an industrial environment. "Xiao Mo" has been deployed on CATL's power battery pack production line, performing final functional tests before shipment. To date, it has completed over 1,000 battery connector insertions, maintaining a success rate consistently above 99%, with operational speed approaching that of skilled workers.
"Xiao Mo" has already entered the power battery pack production line.
Embodied intelligence will not soon reach a moment where immediate deployment determines victory or defeat. However, a clearer trend has emerged—competition is no longer just about who has more data, but who can more efficiently acquire real-world scenario data and build a higher-frequency data-model flywheel loop.
After achieving a phased valuation leap, Qianxun Intelligence will simultaneously invest in the model's generalization capabilities and continue to expand its data scale advantage, accelerating model iteration through high-frequency feedback from the real world.
Looking back at GPT-2 in 2019, it may seem insignificant, but as scale continued to grow, the returns from generalization capabilities rapidly multiplied. Now, the same inflection point is being replayed in the field of robotics.
This article is from the WeChat public account "Machine Heart" (ID: almosthuman2014), authored by Sia.