Amap Releases ABot-Earth 0.7, a 3D-Native City World Model, as Its Store Ranking Goes Fully AI

Amap, the Chinese mapping platform known at home as Gaode, said on September 10 that its year-old store ranking feature is going fully AI. The 2026 edition uses AI algorithms to weigh different navigation and in-store visit behaviors, so that stores that hold up to repeat scrutiny surface more reliably.

The upgrade rests on Amap's spatial-intelligence stack and its core engine, ABot-Earth 0.7, which the team calls the world's first 3D-native city world model. According to Amap, a single satellite image or a short text prompt is enough for it to generate a 3D city with realistic streetscapes on a consumer GPU in roughly ten minutes — which the company says is 1,000 times more efficient than traditional methods.

Coverage is the most visible result. By the team's account, the model already spans more than 190 countries and 300 cities, forming the 3D map that covers the largest number of countries and regions. Google Earth, by contrast, offers 3D scenes only for the core of a few major cities, leaving most smaller and less developed areas in flat satellite imagery.

On technique, Amap chose a 3D-native route. Classical 3D reconstruction captures imagery, computes building positions and shapes, and finally restores surface textures; extending coverage means continuous investment, with a single update cycle running months or even years. ABot-Earth 0.7 instead learns and generates 3D Gaussian splats directly, so the output carries 3D structure by construction. To cope with the millions of splats in a single urban block, Amap designed a native 3DGS compression-and-reconstruction scheme that compresses a complex scene into latent space, generates there, and decompresses back into 3D Gaussians.

For rendering, Amap built a native 3DGS multi-level decoder (LOD) that produces hierarchical Gaussian structures directly without costly post-processing, enabling real-time browsing, while a sliding-window inference mechanism blends overlapping regions intelligently to remove visible seams. Amap says the approach cut hole and breakage rates by 97 percent and can generate more than 10 square kilometers of 3D scene in a single inference pass. The company also acknowledges limits: satellite imagery cannot see building facades, so the model has to infer them, and famous landmarks are handled by a hybrid pipeline that pairs world-model-generated districts with separately modeled landmarks stitched together by an agent.

CEO Guo Ning breaks the model's understanding of the real world into four layers: a 2D network topology describing how roads, buildings and places connect; 3D space showing physical structure; time variation that lets the same space appear different at different moments; and real-world flows such as foot and vehicle traffic. Stacked together, they form what Amap calls spatial intelligence. Against Fei-Fei Li's taxonomy of world models as renderers, simulators and planners, Amap says it is one of the few companies to have built all three inside a single architecture, sharing a causal Transformer underneath.

On the product side, since committing to a broad AI push in 2025, Amap has shipped the open-environment autonomous embodied robot Gaode Tutu, a city-scale simulation training ground for embodied AI, traffic-light countdowns, the Eagle Eye hazard-warning feature, and the fully AI-upgraded 2026 store ranking. Amap says its spatial-intelligence models and research, including the ABot family, have taken more than 20 first-place finishes and state-of-the-art results worldwide, while insisting that leaderboard scores are table stakes and real user adoption is the real test.

From store rankings to 3D city generation, Amap's logic is to turn two decades of spatiotemporal data into a spatial-intelligence platform that can generate, simulate and edit. The open questions are whether the claimed 1,000x efficiency gain holds up across many more cities and at larger scale, and whether this stack can really become the infrastructure layer Amap describes between robots and the physical world.