Quant Group Validates Its Physical AI Foundation Model in a Working Restaurant Kitchen

At its GTC conference, NVIDIA shifted the industry’s focus from “selling GPUs” to “selling AI factories” — a signal that the global AI industry is moving decisively from digital interfaces into the physical world. Capital markets are repricing accordingly: Field AI, which does not manufacture its own robots, is reported to have been valued in the region of US$10 billion. What the industry lacks is not hardware, but the capability layer that makes that hardware work out of the box in real-world settings.

That capability layer is what Quant Group is targeting. The company does not manufacture robot hardware; it is developing an open physical-world foundation model that generalizes across scenarios and robot platforms. Its growth potential depends on how well the model generalizes, rather than on hardware production capacity or the ability to delivery project.

The Endgame in Physical AI: A Cross-Scenario Foundation Model for the Physical World

Physical AI is the ultimate frontier of artificial intelligence,with companies competing to sell hardware, solutions or AI capabilities. The first two business models scale project by project, placing clear limits on growth. The third is built around a foundation model— “train it once, deploy across multiple scenarios”,with marginal costs approaching zero.

Quant Group has chosen the third path. Its architecture separates software from hardware: the physical-world foundation model is not tied to any single hardware platform, and the same underlying model can run on different hardware platforms. The test is whether a single underlying model can transfer learned skills across different tasks, different materials and different types of physical disturbance.

Real-World Trials: Repeated Closed-Loop Validation in a Live Kitchen

Quant Group is one of a small number of companies with a Physical AI focus listed on the Main Board of The Stock Exchange of Hong Kong Limited. It has recently released several embodied-AI demonstrations for commercial kitchens, each validated through repeated testing and deployed under real, dynamically changing kitchen conditions. From the mid-2026 demonstrations — sandwich preparation using dexterous manipulation, shopping-bag sorting, steak seasoning and coordinated dual-arm milk-tea preparation— to the four most recent ones, the company’s embodied-AI capabilities have progressed from dexterous manipulation to handling a combination of materials, another step toward a full Physical AI foundation model.

Preparing hand-pounded lemon tea tests closed-loop control in a precision manipulation task. With the left hand steadying the cup and the right hand pounding repeatedly, the “one arm moving, one arm steady” dual-arm coordination is underpinned by an RSI self-evolving loop that continuously records trajectories, identifies deviations and adjusts rhythm during drink preparing.

Demo image: hand-pounded lemon tea — live commercial-kitchen test

Preparing flaky flatbread tests the ability to track task progress over a long sequence of actions. Twenty-two consecutive actions require the robot to remember which step it has reached and which tool comes next: System 2 handles global planning and exceptions while System 1 executes actions in real time. This demonstrates the agent’s procedural memory and coordinated use of different tools in an unstructured environment.

Demo image: flaky flatbread  — live commercial-kitchen test

Preparing stir-fried shredded potatoes tests the robot’s ability to respond to the changing conditions of Chinese wok cooking. The slippery shreds can easily slip during transfer, the materials shift continuously in the wok, and dual-arm coordination demands millisecond-level timing — one hand working the wok, the other holding the spatula.The task involves relatively few steps, but the robot must respond to frequent disturbances during physical interaction.

Demo image: stir-fried shredded potatoes — live commercial-kitchen test

Preparing egg cake tests whether the model can generalize across multiple types of materials. High-viscosity batter requires anticipating how the fluid will flow; unevenly sized pieces of crushed peanut require precise, localized sprinkling; folding the soft cake demands extremely precise and continuous adjustment of contact force. The task tests the foundation model’s ability to transfer learned skills when the object changes.

Demo image: egg cake — live commercial-kitchen test

Following these kitchen trials, the next step is cross-scenario reuse — extending these capablities beyond food service to a broader range of hardware platforms.

The Moat Behind “Selling AI Capability” Is a Vast Amount of  User Data, Not Parameters

The key difference between Physical AI and language models is that language data can be harvested from the internet in effectively unlimited quantities, whereas physical-interaction data can only be generated through operation in the real world. Quant Group runs its trials in a restaurant kitchen rather than a laboratory precisely because the kitchen’s complex, unstructured conditions cannot be fully reproduced in simulation. Each real-world trial strengthens this data advantage: a data set of genuine physical interactions that can only be built up over time.

Quant Group’s consumer e-commerce business provides funding for this development. In 2025 the company reported revenue of RMB 1.035 billion and net profit of RMB 195 million, while gross merchandise value (GMV) on its Yangxiaomie platform exceeded RMB 10 billion. 30 percent of the proceeds from its July placing were allocated to research and development of the Physical AI foundation model, and a further 40 percent to the commercialization of intelligent hardware.

Frost & Sullivan forecasts that Physical AI will grow at a compound annual growth rate (CAGR) of 47.2 percent, with the global market for related industries potentially reaching US$3.25 trillion by 2040. What will determine who secures the “picks-and-shovels” position is not who makes bold claims first, but who first delivers AI capabilities that can be reliably deployed across multiple scenarios, backed by a sufficiently rich dataset of real-world interactions.

Quant Group’s share price has recently shown signs of bottoming out, with its price-to-earnings ratio stabilising at around 19–20 times and its price-to-sales ratio at just around 2 times. These valuations suggest that the market has not yet to fully value Quant Group as a physical AI company. As companies in the sector accelerate their development, more growth stories and investment candidates are likely to emerge, creating new investment opportunities and scope for returns.

About Quant Group (02685.HK)

Quant Group is an AI technology company listed on the Main Board of the Hong Kong Stock Exchange, called Hong Kong’s first listed “Embodied AI” company.The company operates a dual-track business model: e-commerce AI solutions that generate steady cash flow, and physical AI capabilities that extend AI from the digital world into real-world commercial environments including restaurants, retail stores, and automotive showrooms. In 2025, Quant Group reported revenue of RMB 1.035 billion and net profit of RMB 195 million, up 32.6% year-over-year.

IR Contact Information:lR@quantgroup.com