The Yunyu Xingkong large model, built jointly by SenseTime's SenseCore and the Shanghai Municipal Bureau of Planning and Natural Resources, has been comprehensively upgraded, sharpening its specialized spatial reasoning and complex task execution so the model moves from "answering questions" to "doing the work." Building on the upgraded model, the Shanghai bureau unveiled the first batch of government AI agents for the Quantum City initiative.

Yunyu Xingkong: China's First Industry Foundation Model for Planning and Natural Resources

As China's first industry foundation model for the planning and natural resources sector, Yunyu Xingkong is positioned as the intelligent base and hub of the Quantum City model system. It pulls together domain knowledge, spatial data and business rules from urban planning, construction and city governance, and has assembled 13,000 high-quality question-answer pairs and 1.5 billion tokens of training material covering planning and design, surveying and geography, natural resources and other core disciplines.

The upgrade targets two dimensions: knowledge and reasoning.

  • Knowledge: Using OLA (object–relation–action) as a unified organizing framework, the model links text corpora, measurement data and numerical data in a "three-source linkage." Unified object coding, spatial indexing and temporal relations connect text, drawings, vector files, imagery, video, 3D models and real-time sensor data to one another.
  • Reasoning: The focus is on integrating four map-related capabilities — retrieving, identifying, reading and generating drawings — to form a complete chain from data retrieval and spatial recognition to professional judgment and output generation, so the model does not merely "see" a drawing but understands the spatial relationships in it.

First Government AI Agents: Three "Digital Employees"

The first batch of Quantum City government AI agents released by the Shanghai Municipal Bureau of Planning and Natural Resources includes the "Yuzhi Heluo" flood-control agent, the "Better Life in Your Palm" agent and the "Hupai Jiangnan" agent. Three of them — "Yuzhi Heluo," "Eye of the Master Plan" and "Detailed Planning Work Order" — were developed and delivered by the SenseCore team, handling flood-control dispatch, spatial analysis and planning updates respectively.

"Yuzhi Heluo" puts dispatch ahead of the storm. Built for flood safety in the Suzhou Creek basin, it combines weather, water level and drainage flow data to run precise, zone-by-zone and hour-by-hour dispatch. With 48-hour and even two-hour advance assessments plus real-time dispatch simulation, it maintains a dynamic water balance across river network, pipe network and neighborhood storage capacity. SenseTime says deployment has improved spatial computing, dynamic simulation and intelligent dispatch efficiency by 65%, 70% and 80% respectively.

"Eye of the Master Plan" makes the city's vital signs available on demand. It aggregates 18 years of high-quality spatiotemporal data on existing conditions, planning, land title and approvals to build a digital twin base map covering built-up land, farmland, forest, water and wetlands. Data analysis indicators that once took months to develop can now be customized in minutes through natural language.

"Detailed Planning Work Order" lets planning administration happen in conversation. Aimed at the management needs of urban renewal in an era of existing stock rather than new construction, it offers one-click intelligent guidance for corporate applications, a crowdsourcing platform for collective input, and one-click rapid review of planning base maps.

Platform Foundation and the FDE Co-Creation Model

On top of the upgraded Yunyu Xingkong model, SenseCore has built a general-purpose agent platform foundation spanning an end-to-end task chain of model dispatch, information supply, task execution and capability accumulation:

  • Tiered flexible dispatch: Model routing and task decomposition call different models and specialized algorithms according to task complexity. Lightweight models handle high-frequency queries, information extraction and basic validation, while the core large model takes on complex spatial reasoning, scheme assessment and task planning.
  • Data and knowledge foundation: Built on the "three-source linkage," a spatiotemporal data resource system covering knowledge bases, rule bases, case libraries and indicator libraries.
  • Agent development and runtime platform: Integrating task planning, workflow orchestration, context and memory management, model calls, tool calls, multi-agent coordination and runtime monitoring.

The project uses an FDE co-creation model, embedding business experts and engineering teams in frontline operations to form a closed loop of problem definition, data and knowledge preparation, agent development, business validation and feedback-driven iteration.

What Comes Next

SenseTime says it will work with the Shanghai Municipal Bureau of Planning and Natural Resources to keep developing Yunyu Xingkong: moving from an industry-specific large model toward a city-scale macro world model, and from single agents toward coordinated swarms of city agents. It also plans to explore an "open-source city" ecosystem with partners in planning and design, surveying and geography, city governance and data services.