Mobile phone, cockpit, body, China’s largest end-to-end unicorn quietly handed over high scores

Smart Things Author | Jiang Yu Editor | Mo Ying In the exhibition hall of the 2026 World Artificial Intelligence Conference, end-to-side intelligence is almost everywhere. From the active services in the smart cockpit that “understand you better than the co-pilot”, to the privacy protection of mobile phone AI running offline, to the embodied interaction of robots that “understand instructions and perform tasks stably”, on-device intelligence is being intensively implemented on various terminals. If large models want to truly run on terminal hardware such as cars, mobile phones, and robots, on-device AI has become an indispensable "grounding point." However, different terminals have different requirements for AI. Mobile phones, cars and robots all belong to on-device AI, but they are almost three businesses. Hardware manufacturers have successively launched their own on-device AI products, and a series of practical problems have also emerged: whether the model can be adapted to different chips, whether new hardware needs to be redeveloped, whether system upgrades will interrupt existing capabilities, and whether one project accumulation can be transferred to the next device... After on-device AI has truly entered the industrialization stage, the competition is no longer just about model capabilities. What the industry needs are participants who can understand models, chips, operating systems, terminal equipment and delivery processes, and connect the various links that were originally separated. As one of the earliest companies in China to lay out large-scale end-to-end models, implement end-to-end mass production and delivery, and continue to focus on end-to-end AGI, Face-Wall Intelligence has continuously verified the industrial value of end-to-end AI in recent years, and has gradually accumulated cross-terminal implementation capabilities. 今年WAIC期间，面壁智能集中展示了 手机、汽车、机器人、航天 等多个行业案例， 新一代MiniCPM系列端侧模型、全新VLA具身模型、具身智能解决方案，以及”AI制造AI”的最新成果 。 Putting these results together, they all point to one question: Who will organize the scattered chips, models, operating systems and applications to free end-side AI from repeated adaptations? 1. When on-device AI enters industrialization, the difficulty is no longer just the model. Being able to run the model on the terminal is only the first step in the industrialization of on-device AI. After entering the implementation stage, different terminals face completely different engineering constraints.