Megvii Co-founder's AI Startup Raises ¥1B, Bets Against Pure World Models

Megvii Co-Founder's New Venture Secures Over ¥1 Billion

Yuanli Lingji, a one-year-old startup founded by Megvii co-founder and former CTO Tang Wenbin, has secured over 1 billion RMB in financing from investors including Alibaba, Nio, Junlian Capital, and Qiming Venture Partners. This significant capital injection highlights investor confidence in Tang's vision for embodied intelligence, which diverges from prevailing industry trends.

The company has already launched its inaugural embodied native large model, DM0, and entered a strategic partnership with Huaqin Technology to mass-produce its data collection robot, the DOS-W1. This move positions Yuanli Lingji to address the critical data shortage in robotics by enabling high-quality, multi-modal data collection through a distributed, cost-effective strategy.

CEO Argues Unified VLA Models Outperform 'World Model' Hype

Tang Wenbin asserts that a pure "world model" approach, which focuses solely on predicting future video frames to generate training data, is ultimately a dead end. "I myself feel this path is not viable," Tang stated, arguing that if a perfect world model existed, the problem of robot action would already be solved. Instead of relying on simulated data, Yuanli Lingji is developing a unified framework that combines world models with Vision-Language-Action (VLA) models.

This integrated approach aims to create a single model that not only predicts how the world will change but also infers the precise actions required to achieve a desired outcome. This strategy contrasts sharply with competitors who are attempting to generate infinite data from simulations, a method Tang believes cannot capture the nuances of real-world interaction.

Focus Shifts to ROI with 2-3 Year Payback Target

While developing a generalized AI model, Yuanli Lingji is pursuing a highly pragmatic go-to-market strategy focused on specific industrial and logistics applications where traditional automation fails. The company targets complex tasks with high variability, such as packing diverse e-commerce orders, that require human-like dexterity. To win over enterprise clients, the company is disciplined about financial returns.

If a project needs more than five years to pay back, then there is no need to do it. If it is expected to pay back in two to three years, then it's an immediate go.

— Tang Wenbin, Founder & CEO

This focus on ROI is central to its goal of deploying 1,000 sustainable robots per scenario by 2026. To support this, its DOS-W1 robot is designed for reliability and serviceability, featuring modular components that can be swapped in 30 seconds to minimize downtime—a critical pain point in research and industrial settings.