Hangzhou, at the Yunqi Conference, artificial intelligence is accelerating from the digital world to the real world.
On September 23, during the 2026 Yunqi Conference, Banma Intelligent released a new generation of full-mode end-side large-scale model AutoOmni2.0-23B-A3B, and demonstrated the new AutoClaw2.0 smart cabin collaboration service real vehicle solution. At the same time, focusing on the industry trend of artificial intelligence moving from the cloud to the terminal, and from "being able to talk" to "being able to act", Zebra Intelligence has further released a signal to extend towards embodied intelligence.
What this company hopes to explore is not just to make cars smarter, but to use cars, an intelligent terminal that has achieved large-scale industrialization, as a "pilot test site" to continuously train and verify the ability of artificial intelligence to understand users, understand the environment, call services and perform tasks, and extend these capabilities to more embodied intelligent terminals such as robots.
Behind this, a new industrial connection has been formed with the full-stack AI strategy being promoted by Alibaba Group.
From "awakened" to "always present"
In the past few years, smart cockpits have been rapidly popularized, but a real problem has always existed: although cars are equipped with more and more apps and voice functions, users may not necessarily use them frequently.
The reason is not complicated. Driving and riding scenarios naturally require users to reduce screen operations. The traditional "click-open-input-confirm" interaction method is not suitable in cars.
Zebra Intelligence believes that what really needs to change when AI enters cars is not to add more functions, but to change the way people interact with cars.
AutoOmni2.0-23B-A3B, released on September 23, adopts MoE hybrid expert model architecture. According to Zebra Intelligence, in a smart cockpit scenario, the model's ordinary task processing capabilities are comparable to a cloud model with 10 times the parameter amount, and its complex task processing capabilities are equivalent to 80% to 90% of a cloud model with 10 times the parameter amount.
However, Cai Ming, chief product officer of Zebra Intelligence, believes that the most important value of the device-side model is not simply to pursue "faster", "cheaper" or "more privacy", but to promote changes in the interaction paradigm.
“The core value of terminal intelligence is to open up a new paradigm of interaction.”
At the Yunqi Conference, Banma Smart demonstrated scenarios such as full-time wake-up-free driving, active smart welcome, and senseless smart parking payment.
This means that AI no longer just waits for the user to say "Hello..." before starting to work, but can continue to exist in the car, respond when the user needs it, and stay quiet when it is not needed.
Cai Ming once used the analogy of a restaurant waiter: in the past, AI was like standing outside a private room, and users called it to come in; while in the future, on-device AI is more like a waiter who is always in the private room, able to observe the scene, understand the context, and proactively provide services at the appropriate time.
From "awakened" to "always present", what actually changes is the relationship between humans and machines.
Why can a car become a "pilot test site" for embodied intelligence?
At present, embodied intelligence has become an important development direction of the artificial intelligence industry.
In the industry display at the 2026 Yunqi Conference, smart cars, embodied intelligence, and smart hardware were placed among the important application scenarios of Physical AI. Alibaba Group CEO Wu Yongming pointed out at the conference that as AI capabilities continue to make breakthroughs, machines are becoming "the main force of thinking" and artificial intelligence is moving further from the digital world to the real world.
In this context, Zebra Intelligence puts forward an industry observation worthy of attention:
Smart cars may be one of the most mature and large-scale "pilot sites" for embodied intelligence.
Compared with robots, which are still in the rapid industrialization stage, cars already have large-scale mass production capabilities, mature supply chains and real user scenarios.
A smart car involves multi-chip platforms, operating systems, sensors, voice, vision, navigation, car control, multimedia and a large number of third-party services. AI needs to understand people and the environment in a complex physical space, and complete tasks within safe boundaries.
This means that the car itself is a complex "AI laboratory".
Si Luo, chief technology officer of Zebra Intelligence, said that smart cars have formed an installed base of tens of millions and a huge service ecosystem, and can become a platform for large-scale real-scenario verification and continuous iteration of artificial intelligence capabilities.
According to Zebra Intelligence, it has formed an ecosystem consisting of more than 400 service providers.
"Embodiment is the form, and intelligence is the core." Xing Yue said that in the future, the development of embodied intelligence cannot only focus on the movement ability of the robot, but also needs to pay attention to the machine's ability to understand people and the environment.
Under this logic, what kind of body a robot has is important, but what determines whether it can truly enter complex scenes such as homes, offices, and commercial services is whether it has strong enough "understanding capabilities."
The intelligent development of automobiles in the past decade has just provided a realistic environment that already exists on a large scale.
Extending from the “brain” of the car to more terminals
Zebra Intelligence’s technology route is also evolving from smart cockpits to broader personal intelligence.
According to reports, Yuanshen AI is moving from smart cabin AI to Personal AI, and is promoting connections at three levels: connecting the computing power and services of the cloud and the car, connecting multiple terminals of users, and connecting users and services.
This means that the car is no longer just an independent terminal for AI applications, but may become an important node in the user's personal intelligence system.
The ecological intelligent agent solution jointly demonstrated by Zebra Intelligence and AI data terminal company YoooClaw attempts to open up the context of users' "1 hour in the car" and "23 hours outside the car" under the premise of data compliance, so that AI can gradually evolve from a single device assistant to a cross-terminal personal agent.
From this perspective, the meaning of cars is changing.
In the past, a car was a means of transportation; later, it became an intelligent terminal; in the future, it may become a "personal AI portal" that can understand users, call services, and connect different smart devices.
This also provides a technical foundation for Zebra Intelligence to further explore the field of embodied intelligence.
Alibaba’s AI ecosystem is moving from the cloud to more real-life scenarios
This technological evolution path of Zebra Intelligence also echoes Alibaba Group’s AI layout in recent years.
From large models, cloud computing to agents, to smart cars and embodied intelligence, Alibaba is pushing AI further from model capabilities to real industrial scenarios.
In the exhibition setting of the 2026 Yunqi Conference, Alibaba places Agent applications, smart cars, embodied intelligence and smart hardware in important positions in the implementation of the AI industry, demonstrating the industry trend of AI evolving from "conversation" to "service" and from single-point tools to system capabilities.
For Banma Intelligence, cars are an important terminal connecting this industrial chain.
From AliOS to smart cockpit, from Yuanshen AI to AutoOmni, and then to AutoClaw, Zebra Intelligence continues to explore the same problem: how to let AI truly enter users' daily lives and play a role in complex real-life scenarios.
If the large model solves the "cognitive ability" of artificial intelligence, then the agent solves the "action ability", and embodied intelligence further requires AI to have the ability to enter the physical world, understand the physical world, and interact with the physical world.
Cars are already leading the way in this process.
The AI business model may also shift from “selling functions” to “operating users”
After AI enters cars, it’s not just the products that change.
Zebra Intelligence is also exploring new business models.
Business relationships in the traditional automobile industry start with vehicle sales and extend to after-sales services. But as cars become smart terminals, "buying a car" may no longer mean the end of the user relationship, but the beginning of the user relationship.
According to Banma Intelligence, about 20% of the company's revenue currently comes from subsequent service operations and content operations, including third-party application payment commissions, and profit sharing with car companies around future services and content. There is still room for further improvement in this ratio in the future.
In this model, AI becomes a new portal connecting users and services.
In the past, users needed to open several apps to complete tasks such as ordering food, navigation, parking, and entertainment; in the future, users may only need to express their needs, and AI will understand the intention and complete cross-application and cross-service task organization.
This is why Zebra Intelligence believes that end-side AI has real commercial value.
Because after the interaction method changes, the distance between users and services will also change.
In a sense, AI first changes "how people and machines talk", then changes "how machines understand people", and ultimately may change "how machines complete things for people".
Starting from a car, let AI have more "bodies"
Standing at the 2026 Yunqi Conference, an important window for observing changes in the artificial intelligence industry, Zebra Intelligence is showing a relatively clear technological evolution path:
From the operating system to the smart cockpit; from voice interaction to multi-modal AI; from "awakened" to "always present"; from a single car terminal to Personal AI; and from the mature smart terminal of the car, to further explore the extension to embodied smart carriers such as robots.
The core behind this path is not to simply copy automotive technology to robots, but to accumulate the user understanding, terminal-side AI, multi-modal interaction, agent and service ecological capabilities accumulated in real automotive scenarios over the past decade.
For embodied intelligence, the “body” determines what the machine can do, while the “brain” determines whether it knows why it does it, for whom it does it, and when it does it.
Therefore, when robots move from the laboratory to real life, what the industry really needs to solve may not only be to make the machines "move", but also to make the machines gradually "understand people."
From this perspective, smart cars have provided a complex pilot test site with tens of millions of real users participating.
The next step that Zebra Intelligence hopes to explore is to bring the AI capabilities that have been proven for a long time in cars to more terminals, so that artificial intelligence can move from "answering questions" to "being able to understand, make decisions and act."
This may be another possibility for smart cars to connect embodied intelligence.
Starting from a car, AI is trying to have more "bodies" and is also redefining the way people and machines get along.
