FAW's Ben Xu: From Feature Stacking to Proactive Coordination in AI-Defined Vehicles

Gasgoo Munich-On September 18, at Gasgoo's 4th AI-Defined Vehicle Forum 2026, Ben Xu, Chief Engineer of E/E Architecture, FAW R&D General Institute, delivered a speech titled "Smart Vehicle Technology Practice Sharing." Centering on the exploration and practice of multi-domain intelligent agents for entire vehicles, Xu systematically outlined FAW's assessment of industry trends, the strategy for building a unified AI model system, and the roadmap for implementation.

Xu divides the automotive industry's evolution into three progressive stages. Before 2020 was the era of functional stacking, centered on hardware configuration with deep software-hardware coupling. From 2020 to 2026 marks the software-defined vehicle era, where domain control architectures and OTA enable rapid iteration. Post-2026, the industry enters the AI-defined car era, characterized by unified large models and cloud-vehicle integration to achieve cross-domain active intelligence. He notes that an AI Car is a data-driven, self-evolving intelligent agent with three core traits: data-driven decision-making, full-cycle OTA evolution, and a complete perception-decision-execution loop. The ultimate goal is to build a new ecosystem of intelligent mobility featuring human-vehicle symbiosis. "This is not just a transformation of technology form, but a fundamental reshaping of product organization and user experience," Xu stated. "The focus of competition will ultimately center on the level of AI intelligence."

Addressing industry pain points, Xu highlighted six intertwined challenges facing the transition to the AI Car stage. These include fragmented models leading to redundant development and data breakpoints creating silos. Others are functional verification reducing module reusability and a lack of evaluation standards to quantify safety. Finally, hardware mismatches cause scattered computing power and algorithm needs. He emphasized that AI has become a core strategy for automakers for three reasons. In terms of market competition, the decisive factor in the industry's second half has shifted to the reasoning and decision-making capabilities of AI models. Regarding user value, AI transforms vehicles into intelligent partners that understand their users, reshaping the mobility experience. On the technology front, cars are accelerating into the stage of embodied intelligence.

Regarding the technical path, Xu outlined the evolution logic of AI models. It starts with small models focused on unidirectional recognition and control. It then moves to fusion models connecting perception, prediction, and planning, though these remain siloed. Finally, it arrives at a unified model platform stage, pursuing a unified runtime environment and Agent orchestration. He stressed that a unified large model does not mean a single model covering all scenarios. Instead, the core is a unified runtime framework, data standards, and an Agent system. This maintains a specialized division of labor among models. The central concept is "unified model platform + model matrix framework + Agent." In terms of model application, six core capabilities are required: intent-driven operation, runtime management, edge-cloud scheduling, hierarchical memory, a controllable and traceable permission sandbox mechanism, and observability and resilience. Safety is treated as a primary scheduling and governance object.

On implementation, Xu explained that FAW has chosen a path of "rapidly developing products for vehicle launch first, then gradually releasing system capabilities." The approach uses a complete intelligent loop coordinated by a platform-based model matrix. The cloud handles data training and OTA distribution. Meanwhile, the vehicle side boasts heterogeneous computing power of over 1,000 TOPS. This is capable of supporting large models with 10 billion parameters. Currently, the industry has formed three mainstream implementation routes: full-stack self-development, platform cooperation, and domain-strong collaboration. Automakers can adapt their choices based on their own technical foundations. "The key to our current competition lies not elsewhere, but in who can first get the shared products and the foundational platform up and running," Xu noted.

Xu concluded by stating that FAW will embrace the AI revolution, explore the path of multi-domain AI agents for entire vehicles, and build smarter, safer cars to continuously upgrade the user mobility experience.