Spirit AI Secures 2 Billion in Funding and Joins the Unicorn Club in Just Two Years

Spirit

Qianxun Intelligent (Spirit AI), a company specializing in embodied intelligence, has recently completed two rounds of financing totaling nearly 2 billion yuan, resulting in a valuation exceeding 10 billion yuan. The financing lineup features significant players from the industry, including Yunfeng Fund, a leading venture capital firm, and various top-tier state-owned institutions. Notable investments also come from Chao Shun Investment, TCL Venture Capital, and Minghui Investment, among others, showcasing a comprehensive support structure from top capital firms, industry giants, and state-owned entities. Additionally, existing investors such as Shunwei Capital and Prosperity7 have chosen to continue their substantial investments, reflecting their confidence in the company’s technological direction and growth prospects. This funding marks a significant event in the embodied intelligence sector for early 2026.

01 Racing for Funding in the “Unicorn Jungle”

2025 was undeniably a breakout year for the embodied intelligence industry in China, with 329 financing events totaling 39.89 billion yuan, representing a threefold increase year-over-year. The landscape has quickly shifted from a “hundred schools of thought contending” phase to a concentration of leading enterprises, with several companies announcing high valuations exceeding 10 billion yuan in 2026. Qianxun Intelligent’s nearly 2 billion yuan in funding and its valuation above 10 billion yuan have positioned it among the top tier in this competitive market.

02 A “Combining Team” of Industry Veterans and AI Scientists

Founded in early 2024, Qianxun Intelligent has rapidly progressed through multiple funding rounds from seed to Pre-A+, backed by extensive capital investment, largely due to the strength of its team. CEO Han Fengtang is a serial entrepreneur in robotics, previously co-founding RoboRock where he oversaw the delivery of over 20,000 industrial robots across more than 20 industry scenarios. Co-founder Gao Yang, who earned his PhD from UC Berkeley, has significant expertise in computer vision and robotics, having worked with renowned experts like Trevor Darrell and Pieter Abbeel. His research directly contributes to Qianxun’s model capabilities, including the ViLa algorithm and the EfficientZero algorithm, which received high praise from John Schulman, a co-founder of OpenAI.

In 2025, Gao’s team introduced the One-Two VLA architecture, addressing complexities in command execution by incorporating a “fast-slow system” to manage task complexity. This architecture, evolved from prior research, forms the technological foundation for Qianxun’s Spirit series VLA models.

03 “Data Pyramid” and an Unexpected Perspective

Qianxun focuses on the Visual-Language-Action (VLA) technology route but adopts a differentiated data strategy, which Gao refers to as the “Data Pyramid” training concept. Geng Yixuan, a partner at Shunwei Capital, remarked, “Qianxun Intelligent is one of our earliest core projects, and they have consistently chosen a ‘hard but correct’ path.” Gao stated, “Instead of following the traditional ‘world model’ approach that consumes significant computing power, we opted to use vast amounts of human internet video data for pre-training, achieving better results with fewer parameters and significantly reducing computing costs.”

This strategy involves layering data according to information content and acquisition cost, with the base layer comprising massive amounts of low-cost human internet videos, the middle layer featuring interactive data from remote operation and wearable devices, and the top layer consisting of high-precision data generated from real machines. This structure has enabled Qianxun to develop its fifth generation of self-developed wearable data collection devices, reducing costs to one-tenth of traditional methods. To date, the company has amassed over 200,000 hours of diverse real interaction data, with expectations to exceed 1 million hours by 2026. A key insight from Qianxun is that “dirty data is the key to scaling VLA models.” The team discovered that training on diverse, imperfect data yields more favorable scaling curves, emphasizing the value of data diversity over its cleanliness.

This technical route received critical validation in January 2026, with Spirit v1.5 becoming the first domestic open-source embodied model to surpass Pi0.5 in performance. Notably, Spirit v1.5 exhibits zero-shot generalization capabilities, allowing it to perform complex tasks without additional training for new challenges.

04 Practical Verification on Production Lines

While technical specifications and rankings are important, the true test lies in the ability to operate effectively on actual production lines. This gap between demonstration and real-world application poses a significant challenge in the embodied intelligence sector. Wang Tianmiao, a professor at Beijing University of Aeronautics and Astronautics, has highlighted concerns about the industry’s reliance on a limited set of technical paths, suggesting that future competition may shift towards price wars. He argues that the critical breakthrough for embodied intelligence lies in specialized models and data training, emphasizing the need for a complete technological cycle of production capability, data feedback, and business closure.

Currently, Qianxun has deployed robots into the production lines of CATL, where its self-developed Mo robot operates stably on the battery PACK production line, achieving a production success rate above 99% with output matching skilled workers’ performance. The Mo robot has demonstrated a rapid adaptation to production uncertainties and superior flexible operation capabilities. In commercial applications, Qianxun’s Mozi robot has been utilized in JD.com‘s retail environment for interactive demonstrations, exploring the implementation of JD Cloud and Joyinside’s large model in extensive retail networks.

05 How Far Is the “iPhone Moment” for Embodied Intelligence?

As we look back at early 2026, embodied intelligence has evolved beyond mere “future narratives.” Robots are being integrated into factories, warehouses, and production lines, but significant challenges remain before achieving large-scale industrialization. The industry faces three core pain points. Firstly, there is a data bottleneck; the gap between human modalities of language, vision, and action is substantial. Tasks as simple as stacking blocks, which a two-year-old can perform, continue to pose challenges for robots due to insufficient training data and high acquisition costs. Qianxun’s approach, which reduces costs by 90% through self-developed collection devices and its “imperfect data” strategy, is one of the few comprehensive solutions in the industry.

The second challenge is bridging the gap between demonstrations and mass production. While many companies can showcase impressive action videos in labs, maintaining stable operations in real industrial settings presents entirely different challenges. Qianxun has validated its capabilities on CATL’s production lines, but scaling this validation will take time. Lastly, the closure of business models is critical. An investor has pointed out that it may take another 3 to 5 years before embodied intelligence technology is fully integrated into factories and homes, as companies cannot rely indefinitely on primary market financing. The company that successfully navigates the “technological cycle-production capability-data feedback-business closure” loop will have the best chance of surviving in this competitive landscape.

Qianxun’s chosen VLA route is recognized as having the highest potential but also the highest barriers to entry. It seeks to develop a general embodied model that enables robots to comprehend the physical world, rather than merely performing specific tasks. The ultimate goal is profound: when robots truly possess zero-shot generalization abilities, they will transition from being “specialized devices” to “general labor.” Gao has expressed that we are currently at the “Robot GPT-1” stage, potentially reaching stage 3.5 in four years. He candidly acknowledges that for a considerable time, most embodied intelligence will only reach “limited scenario L4,” making widespread general intelligence unrealistic at this stage. This clear-eyed technical assessment may be one reason why investors are willing to back Qianxun, which has not promised an aggressive vision of “entering thousands of households next year” or remained confined to the lab in pursuit of publication numbers. Instead, it has chosen a balanced path: validating its technology on real production lines, using production data to enhance models, and subsequently expanding to more scenarios. The 2 billion yuan in funding provides the ammunition for this long race, but in the increasingly intense “unicorn jungle,” the real test is just beginning.

Original article by NenPower, If reposted, please credit the source: https://nenpower.com/blog/spirit-ai-secures-2-billion-in-funding-and-joins-the-unicorn-club-in-just-two-years/

Like (0)
NenPowerNenPower
Previous February 26, 2026 10:43 pm
Next February 26, 2026 11:48 pm

相关推荐