Bringing AI Voice Large Models to Vehicles

As the wave of automotive intelligence accelerates, more carmakers are choosing CES as the ultimate technology spotlight to unveil their ambitions and latest breakthroughs. The concentration of cutting-edge launches has even earned CES the nickname “America’s second-largest auto show.”

At this year’s event, major technology players including NVIDIA, LG Electronics, and Bosch showcased platforms spanning automotive foundation models and AI-powered cockpits. NVIDIA introduced Alpamayo, a thinking-and-reasoning model slated to debut in the 2025 Mercedes-Benz CLA. BMW, meanwhile, announced its push toward a next-generation vehicle platform, centered on an intelligent personal assistant and an upgraded sixth-generation eDrive technology.

Geely’s Galaxy M9 also made a notable appearance at CES, featuring StepFun’s Step-Audio2 end-to-end voice large model capabilities. Inside the cabin, its English AI assistant—positioned as the industry’s first in-vehicle interaction system built on an end-to-end voice large model—drew crowds and long lines of attendees eager to try it firsthand.

At the Geely booth, visitors queued to experience the Galaxy M9’s intelligent cockpit in action.

StepFun Vice President Li Jing told Sci-Tech Innovation Board Daily that if 2025 marked the “year one” of large models entering cars, then 2026 is likely to see broader deployment of full-stack intelligent cockpit solutions. More notably, he suggested that entry-level “Agent” (intelligent assistant) capabilities could reach scaled mass production and adoption inside vehicle cabins.

End-to-End Voice Large Models Enter the Cabin—A Turning Point for Intelligent Cockpits

Li Auto founder Li Xiang once stated publicly that “AI represents everything about Li Auto’s future.” Today, competitive pressure across the auto market has shifted decisively toward intelligence. Looking ahead to 2026, automotive technology is increasingly centered on redefining the cabin as a “third space.” The car is no longer only a way to get from point A to point B—it’s becoming a mobile office, an entertainment hub, and an extension of the smart home.

Large-model-powered intelligent cockpits are widely seen as the next major product form factor. Two groups are driving this shift most directly: leading automakers and foundation-model providers. Companies such as ByteDance and StepFun are actively exploring how large models can reshape in-car experiences.

ByteDance’s Doubao large model has already established partnerships with multiple automotive brands, including Mercedes-Benz, SAIC Audi, SAIC Roewe, Changan Mazda, and Dongfeng Aeolus (eπ), among others.

At CES, the StepFun–Geely collaboration model on display—the Geely Galaxy M9—stood out for being ahead on commercialization. The vehicle has been in the market for four months and has reportedly sold 40,000 units in that period.

Li Jing emphasized that as intelligent driving systems mature, the importance of the cockpit will rise even further. StepFun’s end-to-end voice large model, he explained, differs fundamentally from earlier generations of in-car voice assistants: it can interpret emotions, demonstrate stronger conversational “EQ,” retain memory, switch voices dynamically, and act like a natural guide—capable of holding fluid, context-aware conversations across topics, from Korea to Las Vegas.

At CES, international attendees were seen testing the Geely Galaxy M9’s intelligent cockpit interaction.

“Since 2025, large models have been reshaping in-car experiences at remarkable speed,” Li said. “Intelligent assistants can now accept open-ended instructions more naturally, store memories, support multiple languages, and even respond with emotional awareness—evolving from rigid voice tools into true in-car companion assistants. As a third living space, the cockpit’s potential is still far from fully unlocked. It’s a naturally ideal terminal for deploying a ‘super assistant’ Agent product.”

2026: System-Level AI Capabilities Move Into the Spotlight

2025 has widely been viewed as the first year of large models entering vehicles. After a year of rapid iteration, automakers have moved beyond early experimentation and are increasingly exploring deeper integration.

In 2026, the vehicle’s role becomes more defined: the intelligent cockpit must support blended use cases across entertainment, work, and social connection. Natural language interaction is steadily replacing traditional human–machine interfaces.

“In 2025 we mostly saw atomic capabilities being demonstrated,” Li told reporters. “In 2026, we expect more complete cockpit solution packages—entry-level Agents are very likely to appear this year. From personalized memory and proactive execution, to service experiences around content, to an Agent that truly understands you—this full stack should land in clearly identified mass-produced vehicles. And the pace of rollout may be faster than what we see on phones or other terminals.”

Du Fang, Product Director at PATEO, also believes AI will become a native capability of the operating system.

“Many cockpit AI features today still rely heavily on cloud compute,” Du said. “But the real strategic direction is edge–cloud collaboration. Embedding AI deeply into the OS foundation is not only the core base of whole-vehicle intelligence, but also the computing center for integrating experiences across the cockpit, intelligent driving, and chassis domains.”

Du outlined a layered architecture: downward integration into OS-level foundations, and upward expansion into a cloud-based Agent ecosystem. Key components include offline voice AI, on-device vision-language models (VLMs), on-device memory, on-device Agents, as well as cloud orchestration, cloud memory, and cloud Agents.

Edge–Cloud Collaboration Remains a Major Challenge

Despite the momentum, large-model experiences remain uneven, and industry resources continue to tilt more heavily toward intelligent driving. Many automakers have not yet invested enough in cockpit-specific scenario polishing or deeper feature excavation.

Li Jing argued that improving cockpit experiences often requires a fundamental rebuild of existing systems—a step automakers approach cautiously. “Shifting from traditional approaches to large-model-driven experiences demands not only better usability, but also clear proof that the capabilities are reliable,” he said.

At the same time, edge–cloud collaboration has become the main battlefield for intelligent cockpit development. On-device large models can help address limitations of cloud-only approaches—especially in cost, network dependency, and data privacy. “Most models still run in the cloud today,” Li noted. “Designing edge–cloud coordination well is a major challenge, and on-device chip compute directly determines model size and scenario fit.”

Finally, for large models to truly reshape the automotive application ecosystem, the ability to attract more partners and developers will be critical.

“The emergence of Agent-based interaction will significantly change the form of services,” Li said. “It could reshape consumer experience and even business models, altering how services are distributed. Once an assistant becomes truly personalized and understands you, decisions like which flight to book and within what price range could increasingly be delegated to the Agent.”

In the past, the iPhone’s multi-touch interface and app ecosystem redefined user experience, propelled smartphones into the intelligent era, and catalyzed the rise of the mobile internet. Looking forward, large-model-driven Agents—and the operating systems that power them—will also need a mature distribution network and ecosystem to unlock the next stage of automotive intelligence, and to enable entirely new commercial models.