ABot-Earth 0.7: Amap's 3D World Model Turns One Satellite Image Into a City in 10 Minutes

Amap put a specific number on the table on September 10, 2026: one square kilometre of explorable, street-level 3D city, generated in roughly ten minutes, on a single consumer-grade GPU. The model behind that number is ABot-Earth 0.7, unveiled at the company's Street Stars annual gala in Hangzhou and described by Amap — Aliba​ba's mapping and navigation arm — as the world's first native 3D urban world model. It succeeds ABot-Earth 0.5, which shipped in June, and unlike most research releases of this shape it is already in production inside Amap's own consumer app, where the company says more than 880 million people have used the Street Stars feature it now drives.

That last part is what makes this more than a demo reel. A world model that exists only as a paper and a project page is a promise. A world model rendering scenes inside an app that hundreds of millions of people open to decide where to eat is a product. ABot-Earth 0.7 is the second one, and the interesting questions are about what that means for everyone else, not about whether the launch happened.

What actually shipped

The mechanism is the part worth understanding, because it is genuinely different from how digital Earth products have historically been built. Goog​le Earth and its predecessors were assembled by capturing reality — aerial survey flights, satellite passes, photogrammetry, point-cloud reconstruction — and stitching the results into a browsable mosaic. The result is a high-fidelity archive of places at the moment they were photographed, and it can tell you nothing about a place that was never flown over at sufficient resolution.

ABot-Earth takes a native 3D path instead. It is trained on spatiotemporal data and generates 3D Gaussian Splatting (3DGS) urban scenes end-to-end from a geospatially referenced satellite image — or, in the interface Amap demonstrated, from a text prompt. Because the scene is generated rather than reconstructed, Amap says it can autonomously complete and extend a space rather than being limited to predefined routes, and it maintains spatial consistency as a viewer moves from a city-wide overview down to landmark detail. The company also states the output is editable assets compatible with Unreal Engine and Unity, which matters more than it sounds: it means the generated environment drops into existing simulation and game pipelines instead of living only inside Amap's viewer.

Amap's headline claims for 0.7, all vendor-reported and none of them independently benchmarked:

• Speed — roughly 10 minutes to generate a kilometre-scale 3D urban scene, on one consumer-grade GPU.

• Efficiency — up to 1,000x faster than traditional 3D reconstruction methods, at about 1/100th the cost, per Amap's own comparison.

• Coverage — more than 196 countries and regions, up from the 190-plus Amap claimed for ABot-Earth 0.5.

• Modality — satellite imagery or text prompt as input; Amap also describes the model as fully multimodal and predictive.

• Output — 3DGS scenes with hierarchical level-of-detail structures, exportable to Unreal Engine and Unity.

What six months bought

ABot-Earth 0.5 arrived on June 8, 2026 with a technical report on arXiv and a companion repository under Amap's computer-vision lab. That release established the core claim — satellite imagery to seamless 3DGS environments at under ten minutes per square kilometre on a single consumer GPU, with hierarchical level-of-detail for real-time web visualization — and it covered 190-plus countries. Reading the two releases side by side, 0.7's advances are less about the headline generation speed, which 0.5 already advertised, than about the tier above it.

• World-model framing — 0.5 was presented as a generative 3D reconstruction system. 0.7 is presented as a world model with a prediction layer, which is a different claim about what the thing is for.

• Coverage — 190-plus countries to 196-plus, a modest expansion rather than a step change.

• Productization — 0.5 pointed at a beta site and a whitelist. 0.7 ships inside Flying Street View 2.0 and the Street Stars app surface.

• Export targets — explicit Unreal Engine and Unity compatibility, which 0.5's report did not foreground.

• Code — worth noting that 0.5's GitHub repository was described as hosting the technical report and academic discussion but no implementation code. Nothing published so far suggests 0.7 changes that.

Flying Street View 2.0 is the part with users

The consumer-facing expression of all this is Flying Street View 2.0, which replaces the fixed-route flythrough of the original with freely explorable 3D space. Amap's own examples are concrete: preview the actual sightline from a specific theater seat before buying the ticket, navigate the interior of a shopping mall, judge the terrain of a scenic area before committing to the trip. Users adjust position, angle and height with a joystick, and the feature covers complex buildings and large scenic areas rather than just street corridors. Trials are live for selected venues in cities including Beijing and Hangzhou, reachable through the Street Stars section of the app.

The scale Amap is operating at is the reason this is not a niche tooling story. Street Stars launched in September 2025 as an AI-driven local-discovery ranking; Amap says more than 880 million users have used it and collectively travelled 36.6 billion kilometres through it. The company also claims the ranking has commercial teeth — businesses in its 2025 "Local Favorites" list saw average year-on-year order growth of 424%, and "Top Picks" businesses 159%. Those are Amap's figures about Amap's product, and they should be read as marketing arithmetic, but the user count is the load-bearing one: whatever ABot-Earth 0.7 does well or badly, it is doing it in front of an audience the size of a large country's population.

Why "not Goog​le Earth" is the honest frame

Amap's positioning is deliberately not "a better globe." The company's chief executive, Guo Ning, framed it as a category distinction: "If large language models help us process text, I hope Amap's spatial intelligence can help us better explore the world." The architecture he described has three layers — 3D spatial representation (the structure of the world, from planet to interior), dynamic perception (real-time change such as traffic, crowds and conditions), and spatiotemporal inference (predicting what happens next).

That third layer is where the world-model claim lives, and it is why the comparison to a static archive undersells what Amap is attempting. A reconstruction shows you a place. A world model that can be queried and rolled forward is something you can plan against — and the applications Amap is pointing at are explicitly the physical-AI ones: simulation training grounds for embodied robots, digital-twin systems for cities, synthetic data for autonomous driving, and low-altitude flight. The company is also pairing the model with a conversational layer called Navigation Live, which reads nearby buildings, attractions and businesses through a phone camera and answers questions about them in natural conversation.

The data position underneath all of this is the part competitors cannot quickly copy. Amap cites nearly one billion monthly active users, Beidou positioning peak daily calls approaching one trillion, and decades of accumulated spatiotemporal data covering road networks, points of interest and live traffic. General-purpose model labs do not have that, and it is the real moat — not the 3DGS architecture, which is published work.

What Amap has not published

This is where the launch coverage has been thinnest, and it matters for anyone trying to decide whether ABot-Earth 0.7 belongs in a plan.

There is no public API. The experience site at abot-earth.amap.com is live, but generation is invite-only — the interface describes the creation feature as in beta and offers a whitelist application, and the account behind it showed a zero credit balance. There is no published pricing, no rate card, and no developer documentation describing how an external system would call the model. The site is also Simplified Chinese throughout, with no locale switcher and no English version, so evaluating it in English currently means working from Amap's press materials.

There are no model weights and no model card. The 0.5 release put a technical report on arXiv and a repository on GitHub, and that repository was explicitly a home for the report rather than an implementation. Nothing announced with 0.7 indicates an open-weights release, and Amap has not said otherwise.

Most importantly, there is no independent benchmark. Every performance figure in this article — ten minutes per square kilometre, 1,000x efficiency, 1/100th cost, 196-plus countries — originates with Amap. That is not a reason to disbelieve them; published 3DGS research from the same lab has been peer-visible, and the consumer rollout is itself a form of proof that the system runs at scale. It is a reason to hold them as claims rather than measurements until someone outside Amap reproduces them on hardware they own. "World's first native 3D urban world model" is likewise Amap's framing, and the boundaries of that category are Amap's to draw.

What this means if you build with models

ABot-Earth 0.7 is not a language model and it is not something you can route through an API today — OrcaRouter does not serve it, and no one outside Amap's whitelist can call it. Being clear about that is the point, because the practical takeaway for most teams is narrower and more useful than the launch headlines suggest.

What a spatial application actually needs is two different things: a world model that generates or reconstructs the environment, and a language model that reasons about it in conversation. Amap's own stack reflects that split — ABot-Earth renders the place, and a separate conversational layer handles Navigation Live's questions about it. If you are building the second half of that pattern, the language side is ordinary routed-model work, and it is available right now: one API across 200-plus models at provider list price with 0% markup, automatic failover when a provider degrades, a routing DSL for composing several models into one call, and model fusion when you want a panel answering together. That is the layer you can build against today while ABot-Earth 0.7 stays behind a whitelist.

What would change the read

Four things are worth watching, and each would move ABot-Earth 0.7 from an impressive vendor announcement to something a technical team can plan around. First, an independent evaluation — anyone reproducing the ten-minutes-per-square-kilometre claim on their own single GPU turns a marketing number into a reference point. Second, a published API and price, which is what would make the model addressable rather than merely visible. Third, whether the 0.7 technical report appears at all; 0.5 had one within days, and its absence six months later would be its own signal about how much of 0.7 is architectural rather than incremental. Fourth, whether the open-weights question gets an answer, since a 3DGS world model with downloadable weights would be a materially different proposition for the embodied-AI and simulation teams Amap says it is courting.

Until then the honest summary is this: Amap has shipped the most productized 3D world model anyone has put in front of consumers, at a coverage and cost profile that is genuinely unusual, and it has done so without publishing the numbers, the weights, or the interface that would let outsiders check the work. Both halves of that sentence are true, and the second one is not a criticism of the first — it is simply what the evidence currently supports.