上海空间智能基座大模型顶层架构与建设思路OACHSSCD

General large language models have substantially improved semantic understanding and text generation,but they still struggle to represent geometric relations,spatio-temporal evolution,and domain-specific rules in urban space robustly.Against the backdrop of Shanghai's Quantum City initiative,this study addresses the needs of megacity spatial planning and natural resource governance and proposes the technical architecture and development pathway of Yunyu Xingkong,a spatial intelligence foundation model.The model adopts a full-stack pluggable architecture comprising one domain foundation model,six vertical models,and N scenario-specific agents.It uses Taige to establish a city-scale three-dimensional spatial reference spanning airspace,the surface,and the subsurface;organizes model-accessible governance semantic units through spatial Miliao;and integrates four-dimensional spatio-temporal encoding,knowledge embedding,and staged training to extend conventional large language models with spatial understanding,spatial reasoning,and spatial generation capabilities for planning tasks.Quantitative validation was conducted through multimodal recognition of professional planning drawings.The evaluation set comprised nine planning drawings and 714 questions,of which 705 text-recognition,graphic-recognition,and association-matching questions had complete cross-model results.Overall accuracy was 18.72%(132/705)for Doubao Seed2.1,53.19%(375/705)for Yunyu Xingkong before training,and 81.70%(576/705)after training.Post-training accuracy increased by 28.51 percentage points over the pre-training model and exceeded Doubao Seed2.1 by 62.98 percentage points.Post-training accuracy reached 93.06%,78.75%,and 79.41%for text recognition,graphic recognition,and association matching,respectively,although graphic recognition remained lower than the other two task categories.Under the current evaluation conditions,the results indicate that the proposed pathway improves multimodal recognition and relational inference for planning drawings,while robust recognition of complex graphic elements requires further validation.This study provides a verifiable technical framework for applying spatial intelligence in spatial planning and natural resource governance and delineates its current scope of applicability.