Qingdao Port, part of Shandong Port Group, has rolled out its port-industry large model "Xinggang" across production, safety and service operations, Science and Technology Daily reported. The port's container stowage agent now completes stowage planning for a 24,000-TEU ultra-large container vessel in under 5 minutes — work that took experienced planners 3 to 5 hours — a 263-fold efficiency gain over manual work. The model has been rated as reaching an advanced international level in port operations and management by the data security institute of the National Industrial Information Security Development Research Center.
Dispatching comes first. A full berthing plan must weigh more than 180 business rules and 132 influencing factors; a single schedule change once triggered fleet-wide rework. The full-element dispatch agent now weighs seven dimensions of factors and proposes optimized plans automatically — planners need just two steps, and a sudden change can be re-planned in about 90 seconds.
Stowage constraints are denser still: seating tens of thousands of containers must satisfy over 300 hard constraints spanning vessel stability, yard inventory, loading sequence and re-handle control. The agent absorbed over 1,000 port-and-shipping business rules and tens of millions of historical operation samples, refining through 500 rounds of autonomous iteration.
Safety oversight has been rebuilt the same way. A port-wide safety agent patrols preset routes, recognizes targets on camera and catches violations within seconds with instant alerts. It covers 50 complex scenarios across four high-risk areas — equipment maintenance, human-machine interaction, work at height and road traffic — backed by 133 algorithm models; supervision coverage reached 95% and staff violations fell 32%.
Architecturally, Xinggang layers a general foundation model with port-industry models, vertical-domain models and agent applications. According to Qingdao Port, adapting the system to another port requires no re-architecture of the underlying framework — only adjustments to business rules — and the field-tested methodology has been packaged as replicable experience from the port's role as a transportation-sector AI pilot-testing base.
Ports are rule-dense, low-tolerance environments; the value of large models shows not in conversation but in wrapping hard-constraint problems like dispatch and stowage into executable agents. Xinggang's path offers a reference for similar process-heavy industries.