Spirit AI千寻智能
LatestSeries A+ (Jun 3, 2026, RMB 1.5B)
over RMB 20B (breakthrough in Apr 2026)
Spirit AI (Hangzhou) Technology Co., Ltd.
Focused on building a 'general brain' for robots; Spirit v1.6 topped RoboArena as global No. 1; nearly RMB 5B across 4 rounds in 3 months; Jack Ma (Yunfeng Fund) and Lei Jun (Shunwei Capital) teamed up on a rare joint bet; 'model-first' strategy
数据截至2026-08-19
About
Spirit AI was founded in Hangzhou in January 2024 by Han Fengtao, former co-founder and CTO of Rokae Robotics, with Tsinghua assistant professor Gao Yang (one of the 'four returnees from Berkeley') and Zheng Lingyin as co-founders, focused on building a 'general brain' for robots. Unlike the industry's hardware-first thinking, it puts 80% of resources into models and data, adopting a VLA + world model fusion architecture; by early 2026 it had accumulated over 200,000 hours of real interaction data (expected to pass 1 million hours in 2026) and built data-collection networks in over 100 Chinese cities. Spirit v1.5 was the first Chinese open-source model to surpass Pi0.5, and Spirit v1.6 topped RoboArena as global No. 1 in June 2026. On commercialization, Moz robots landed on CATL's world-first scaled humanoid power-battery PACK line (99%+ task success rate), with partnerships also with JD.com, Bosch and TCL. Between February and June 2026 it closed four rounds totaling nearly RMB 5B — Lei Jun (Shunwei) and Jack Ma (Yunfeng) jointly taking major stakes in the same embodied intelligence company for the first time — with valuation breaking RMB 20B in April, aiming for the global top 3 in embodied brains.
Technical Approach
'Model-first' strategy: put 80% of resources into models and data, training a general foundation model on massive real data so that one model adapts to multi-brand, multi-spec hardware — jumping out of the component price war. It adopts a VLA + world model fusion architecture, iterating via a 'data pyramid' (internet video, real-scenario interaction, reinforcement learning data) and a 'dirty data' training philosophy; after pre-training, only 1 hour of fine-tuning data is needed to reach a 95% task success rate.
Commercialization(6)
- CATL: the Zhongzhou base officially entered operation in December 2025 — the world's first power-battery PACK line with scaled humanoid embodied robot deployment, with Moz robots on duty at a task success rate above 99%
- Bosch Group: strategic cooperation agreement signed in May 2026, covering data collection and model training, industrial scenario deployment, and core component supply
- JD.com: coffee-making, shopping-guide and other scenarios at JD MALL, with Moz robots already in service
- TCL: industrial capital plus application testing
- Xiaomi/Huawei: industrial capital investment
- Note: partnerships with CATL, Bosch and others remain benchmark pilot projects for now; large recurring purchase orders have not yet landed (per National Business Daily, 2026-06-03)
Position & Strengths(5)
- Technical summit: Spirit v1.6 tops RoboArena globally (surpassing NVIDIA Cosmos3 and Pi0.5), and Spirit v1.5 is the first Chinese open-source model to surpass Pi0.5 — validated by authoritative international evaluation
- Data moat: 200,000+ hours of real interaction data (1 million hours expected in 2026), a 100+ city collection network, industry-leading data density and scenario purity
- Top capital endorsement: nearly RMB 5B across 4 rounds in 3 months; Lei Jun and Jack Ma betting on the same embodied intelligence company together for the first time
- Model-first differentiation: jumping out of the hardware war to focus on the 'general brain', with declining software marginal costs
- Industrial validation: CATL's world-first humanoid production line, plus JD.com, Bosch and TCL partnerships