The integration of AI assistants into vehicles has sparked discussions about who will cover computing and token costs, and how data privacy and security will be ensured.
At the Volcanic Engine booth during the 2026 Beijing International Auto Show, a customer curiously asked staff if the company's AI smart cabin (equipped with the Doubao large model) could sing in the slurred style of Jay Chou. After receiving the instruction, Doubao began singing moments later. "The performance is quite good, it feels like a real person," the staff member remarked while demonstrating the feature.
Activities such as singing, watching dramas, ordering meals, and writing reports—everything that AI agents can accomplish on phones or computers—are now being integrated into vehicles.
Beyond launching new models and competing on hardware, this year's auto show also highlighted a full-scale software competition. The internet sensation "Xiao Long Xia" (Little Lobster) has become a new trendsetter in the automotive industry. From car manufacturers to behind-the-scenes suppliers, efforts are underway to transform voice assistants in smart cabins from merely "listening and speaking" into practical AI agents that can perform tasks.
A notable change at this year's event was that smart cabin solution providers exhibited in the same hall as vehicle manufacturers, signaling their growing industrial significance.
Confronting this wave of AI integration, Lu Fang, Chairman of Voyah Auto, predicted that AI will reconstruct or even颠覆the entire automotive industry. In his view, automobiles are the ideal载体for AI technology, particularly embodied intelligence. Future vehicles will themselves become embodied intelligent products, with AI reshaping product structures, corporate organizations, and supply chain ecosystems, ultimately creating a entirely new industrial landscape. He emphasized that this presents a historic opportunity for Chinese independent brands, which already lead in new energy and intelligent connected vehicle sectors. This advantage provides a technical foundation that overseas competitors currently lack. Lu Fang stated that Chinese companies must deeply integrate large models and autonomous chips into their products while building a new industrial ecosystem.
Lu Fang also revealed that Voyah is internally developing AI large models and applications related to "Little Lobster," with plans for future vehicle integration. "I believe smart cars will become partners to users—not just for searching, but more as assistants," he added.
The 2026 Beijing Auto Show saw numerous mainstream automakers officially unveil their own "lobster" AI assistants, with some already in mass production and others accelerating development. Suppliers are also significantly advancing their efforts.
On April 22, Chinese autonomous driving technology company Horizon Robotics抢先released the vehicle intelligent agent operating system "KaKaClaw," which can parallelly schedule autonomous driving and smart cabin functions via natural language commands. Founder and CEO Yu Kai explained, "KaKaClaw not only has a mouth to talk with you but also has claws to help complete tasks." He described KaKaClaw as an open Agent framework that allows users to choose various backend large models. It understands continuous, ambiguous demands, performs clear task planning and execution, and features a built-in knowledge graph that learns user preferences over time, developing its own personality.
After experiencing KaKaClaw, Peng Bo, Managing Director at Alvarez & Marsal, enthusiastically shared on social media, "KaKaClaw is a DeepSeek moment for the automotive industry—an experience you won't want to leave, with two features: infinite scenarios, universal assistant."
AI company SenseTime's automotive division, SenseAuto, also presented a hardware-software solution compatible with "lobster" agents. Its core product, the debut Sage Box, uses a three-layer architecture—Sage on-device model, Sage OS, and New Member native agents—to create an evolvable onboard autonomous intelligent brain. Additionally, SenseAuto's previously launched AI terminal "Keyou" extends AI capabilities from vehicle cabins to home and office scenarios.
The implementation of multi-agent AI is also a key focus for large model manufacturers, evident by their presence at the show. Volcanic Engine showcased its automotive AI solution based on Agentic AI architecture. Staff informed that the Doubao large model is already deployed in over 7 million vehicles across 50 brands. They highlighted the "Little Lobster" cross-device assistant, which seamlessly switches between car systems, phones, and computers, noting it is one skill within a broader intelligent service ecosystem, with models featuring this technology expected in the third quarter this year. Cooperation models include deep customization with Doubao's IP or using Doubao as a base for automakers to customize names and images.
iFlytek exhibited its Spark next-generation multimodal smart cabin, backed by an open agent ecosystem. While the underlying technology is self-developed, iFlytek collaborates with partners like iQiyi, Amap, Meituan, and Tencent Music for scenario-specific optimizations. Current collaborations include a "drama-watching buddy" with iFlytek, an "AI traveler" with Amap, Meituan外卖integration, and karaoke features with Tencent Music, primarily mass-produced in Chery vehicles.
Smart automotive solution provider Banma Intelligent released its "lobster" integration solution AutoClaw ahead of the show, utilizing a hybrid end-cloud architecture with a 30B parameter MoE model running on-device. At their booth, staff demonstrated the Qwen model's capabilities with AutoClaw, including模糊intent navigation, hotel booking, ticket purchasing, food delivery, and even voiceprint payment. However, they noted that Qwen only entered the automotive sector this year and requires further optimization.
Amid the "lobster" trend, fundamental questions arise: Do users really need AI agents in vehicles compared to phones or computers? Who bears the computing and token costs? How is data privacy and security ensured? Is this a genuine industrial revolution or a new bubble?
SenseAuto CTO Xiao Feng provided a scenario-based perspective, noting that driving environments differ significantly from phone use, with hands and eyes occupied. AI agents can utilize this otherwise "wasted" time for tasks like gathering materials or summarizing reports, activating the value of vehicle space. He described the essence of "lobster" agents as "always on," particularly useful in blurring work-life boundaries where tasks from bosses can be handled during commutes without user intervention.
This transforms vehicles from mere transportation into "mobile intelligent terminals" when AI agents can proactively understand intent, control vehicle functions, and connect external services. A Volcanic Engine staff member illustrated, "I can use Little Lobster for tasks in the car, continue interacting on my phone, or use it on my office computer—this assistant serves us across locations and devices."
However, Xiao Feng emphasized that onboard agent deployment differs from computers or cloud services; in-vehicle scenarios are more easily recognizable, and specialized fine-tuning improves effectiveness and accuracy. "I've always believed cars are an excellent environment for nurturing new technology," he said, asserting that AI cars are a necessity.
For AI companies, full-stack model capabilities from cloud to edge are essential. SenseAuto's early framework development, mass production experience, and integrated cabin and driving capabilities present significant opportunities, Xiao Feng noted.
Despite the potential, computing costs remain a consideration. Yu Kai pointed out that agents like "lobster" cause token usage to surge, while user demands for real-time performance, low cost, better experiences, and privacy protection are urgent. Consequently, the next major trend will be local large model computation, reducing costs, ensuring real-time response, and protecting privacy—representing a huge opportunity.
Reflecting this, several companies introduced edge inference solutions. SenseAuto's SageBox and Banma's AutoClaw offer edge inference, while Horizon claims its Stellar chip can locally run Qwen's 30B parameter model, outperforming the current Mac mini M4 Pro.
Yet, computing and models are just the foundation. Xiao Feng believes future competition will essentially be about "AI + vehicle." Those with deep automotive industry expertise, mass production experience, and full-stack model capabilities will lead. Additionally, an operating system serving as a "glue layer" between models and vehicle systems is crucial. "AI's上限is high, but its下限is uncertain; the key is harnessing it, turning it from a 'pet' into a tame 'ox' that truly serves people," he explained.