How Qianxun Raised ¥3 B in 30 Days: AI‑Powered Robotics Secrets

In late April 2026 Qianxun Intelligent announced a new ¥10 billion financing round led by Shunwei Capital and Yunfeng Fund, with participation from a top RMB fund, Galaxy Yuanhui, Turing Fund, New Ding Capital, Gengxin Capital and others. This was the second large round within 30 days, bringing the company’s total capital raised that month to ¥30 billion.

The round featured a rare joint lead by Lei Jun (Shunwei) and Jack Ma (Yunfeng), signalling strong confidence in the embodied‑intelligence sector, similar to their earlier bets on mobile internet, e‑commerce, smart hardware and cloud computing.

In January 2026 Qianxun open‑sourced the Spirit v1.5 model, a Vision‑Language‑Action (VLA) unified architecture. In public benchmarks it outperformed the then‑leading open‑source model Pi0.5, achieving zero‑shot generalisation across tasks such as wiping, hinge opening and handling flexible objects without additional fine‑tuning.

The company follows a scaling‑law approach—grow model size, feed massive data, and rely on emergent capabilities. It pre‑trains on large internet video corpora, then aligns with real‑world robot interaction data. A 50‑hour robot dataset yields comparable performance gains to a ten‑fold increase in language‑model data.

In a Silicon‑Valley peer test, Generallist AI’s GEN‑1 model raised average task success from 64 % to 99 % and achieved a three‑fold speed increase, using only about one hour of robot data per capability. Qianxun’s own robots demonstrate similar gains with far fewer parameters and lower compute.

Qianxun built a “data‑hands” wearable with three‑finger dexterity and force sensors, turning human hand motions into high‑quality multimodal robot data. The wearable pipeline improved usable data from 30 % to 95 % and cut remote‑operation cost to roughly one‑tenth. To date the team has accumulated over 20 k hours of interaction data; the target is >100 k hours by 2026 and a data‑collection team of 1 k people.

The company entered JD.com’s retail MALL as a coffee‑making robot, collecting expert‑level data while serving customers. It also operates on CATL’s battery‑pack line, completing more than 1 000 insertion tasks with 99 % success and near‑human cycle time. These deployments close the loop: collected data fuels model roll‑out, which in turn improves performance.

Looking ahead, Qianxun will continue to scale model generalisation while expanding its data‑engine capacity. The emerging competitive edge is no longer raw data volume but the efficiency of acquiring high‑frequency, diverse real‑world data and feeding it back into the model flywheel.