On 27 August 2025, the Ministry of Transport said at its regular press conference that China will accelerate the building of an integrated transportation large model. The spokesperson set out the “1+N+X” overall technical architecture - the first time a sector regulator has broken large model construction into a structure that can be divided and reused.
The three layers of ‘1+N+X’
“1” is a shared general technical foundation. It is designed to achieve open compatibility with multiple base models and heterogeneous computing power and to call on them in a concentrated way. The point of this layer is to stop every business unit from building its own foundation: computing power and base-model capability are pooled into a shared layer. See the national integrated computing system.
“N” is a set of domain models trained on that foundation. The MOT proposes bringing in high-quality datasets from specialised fields to improve the ability to solve common industry problems. The value of a domain model lies not in general capability but in the density and quality of industry data. See the high-quality industrial datasets action plan.
“X” is agents serving specific business scenarios. The goal is to form a set of specialised solutions that can be applied directly in industry services and management. Compared with domain models, agents mark a shift from generating content to executing tasks.
Why industry models are co-built rather than built separately
The press conference noted that the MOT had recently guided the establishment of the Transportation Large Model Innovation and Industry Alliance, which has brought together more than 50 leading industry enterprises, top AI companies and relevant universities and institutes. The alliance operates on the logic of “shared computing power, shared data, jointly trained models”.
This path differs from the competitive logic of general-purpose large models. Transportation has scattered scenarios and scattered data, and the return on investment for any single actor is low; concentrating infrastructure and decentralising application development better fits the industry. It is also consistent with the “AI Plus” action’s emphasis on promoting innovation through application. See the State Council’s AI Plus action opinion.
860 scenarios and the first agents
To accelerate deployment, the MOT catalogued 860 typical AI application scenarios and, focusing on areas such as operations monitoring and safety regulation, selected a batch of high-frequency, hard-requirement scenarios. The first agents will be deployed on the National Integrated Transportation Information Platform.
The order - catalogue scenarios first, develop agents second - is worth noting. It treats the scenario list as the entry point for demand rather than building technology first and looking for a use. For companies taking on such projects, the focus of proving capability shifts from model metrics to scenario understanding. See the national standards on agent interconnection.
New compliance questions raised by agents
Agents differ fundamentally from pure content generation: they may autonomously call tools, access data and trigger operations. In transportation, a field touching public safety and public resources, permission boundaries and responsibility allocation need to be redefined.
Specifically, three things need to be settled: which data an agent may call, which actions it may perform, and who is responsible when it errs. Such questions are especially sensitive in public-sector deployments. See the guideline on deploying government AI large models.
Where participants can start
First, identify your position across the “1”, “N” and “X” layers. The compliance focus differs by layer: the foundation layer concerns computing power and data interfaces, the model layer concerns training-data sources, and the application layer concerns scenario responsibility.
Second, set clear ownership terms for data and model assets in alliance cooperation. Shared data presupposes clear rights boundaries.
Third, build security assessment and operation logging into agent design, especially for agents serving regulatory and dispatch scenarios.
How an industry builds its large models shapes how its compliance is organised. When an industry’s AI capability is co-built through an alliance, compliance is no longer a matter for a single company but needs a cross-entity responsibility framework. See China’s AI development and deployment policy.
This article is based on public information from the Ministry of Transport’s regular press conference on 27 August 2025 and is for reference only; it does not constitute legal advice.