Forums at the 2026 Yunqi Conference focus on how Agentic AI is reshaping next-generation terminals, with the industry's AI competition shifting from cloud-based model capabilities to on-device application deployment. Smart vehicles are seen as a crucial testbed and large-scale validation platform for embodied AI. Currently, AI applications in vehicles have achieved widespread hardware adoption, yet user engagement remains low, and conversion rates for intelligent driving paid services are not high, primarily due to the lack of a mature super-intelligent agent interaction model. Zebra Intelligence has introduced AutoOmni, a series of full-modality on-device large models tailored for smart terminals. Its new flagship model, AutoOmni 2.0-23B-A3B, employs a sparse Mixture of Experts (MoE) architecture, achieving cloud-side model performance with nearly 10 times the parameter efficiency in smart cockpit scenarios using just 3 billion parameters. This breakthrough overcomes on-device memory, latency, and concurrency constraints while ensuring privacy protection and cost control. It transforms on-device terminals into local token factories, significantly reducing cloud computing power and token consumption. The model adopts an end-to-cloud collaborative architecture, where the on-device side handles multimodal understanding, planning, and local execution, while the cloud provides knowledge supply and complex reasoning, supported by a three-tier, in-depth on-device AI security system. Relevant technologies have passed multiple international automotive-grade certifications, meeting compliance requirements across different markets. Zebra Intelligence plans to extend this on-device technology solution from cockpits to intelligent driving and further migrate it to various terminals such as smartphones and wearable devices, building a full-scenario on-device AI infrastructure.