Qianjue Technology raises Series A worth hundreds of millions of yuan for predictive world models
What's the deal? Qianjue Technology (千诀科技), a Chinese embodied intelligence startup spun out of Tsinghua University's brain-inspired computing research centre, has closed a series A round worth hundreds of millions of yuan. Jingming CapitalDealroom has a profile for this one. Try Dealroom → led the round, joined by a broad mix of state-backed funds, industrial investors, market-oriented VCs, and family offices. The company will use the capital to build out its proprietary world model architecture, iterate on algorithms, expand its R&D and deployment teams, and deepen commercialisation.
Why now? World models — AI systems that learn to predict how the physical environment will change — are fast becoming the hottest frontier in embodied intelligence. Yann LeCun's foundational work on the concept has spurred a wave of research across generative, causal, and predictive paradigms. The field remains wide open, with no dominant approach yet.
Founded in June 2023, Qianjue bets on a predictive rather than generative approach. Instead of reconstructing future video frames pixel by pixel, its model predicts low-dimensional physical state trajectories — more like how a tennis player tracks a ball's path than how a camera records a scene. The company argues this avoids "feature contamination," where irrelevant visual noise such as lighting and texture degrades the model's understanding of real physics.
It has also developed a distributed prediction architecture inspired by the human brain's modular structure, processing different types of information in separate regions before compressing and predicting. The result, Qianjue claims, is higher sample efficiency and faster inference — needing roughly a tenth of the training data compared to starting from scratch.
Crucially, the company decouples the robot's "brain" (perception, prediction, planning) from its "cerebellum" (motor execution). This lets the same world model transfer across wheeled robots, quadrupeds, humanoids, drones, and cleaning machines without retraining from zero. Qianjue says it has already deployed to around 100,000 terminal devices across hotel cleaning, commercial services, and indoor operations.
What could go wrong? Predictive world models remain an early-stage bet. Long-horizon planning errors can still accumulate, though Qianjue mitigates this with closed-loop feedback — robots predict short sequences, execute, observe, and correct. Real-world deployment also exposes latency sensitivities: even sub-second delays drew complaints from customers, and millisecond-level optimisation proved more impactful than raw model improvements.
Competition is fierce. Multiple Chinese and global teams are pursuing world models for robotics, and no technical paradigm has won yet. Scaling from 100,000 devices to millions while maintaining quality will test both infrastructure and business model.
The signal: Qianjue's classification as a "breakout" company on Dealroom, despite being founded less than two years ago, underscores how quickly embodied intelligence startups can scale when they decouple the AI layer from the hardware. The round's investor mix — spanning state-backed funds, industrial players, and market-oriented VCs — reflects Beijing's coordinated push to industrialise AI in the physical world, with capital flowing towards platform-level "brain" providers rather than single-robot makers.
Read more: 36kr.com
