Genie 4 vs ABot-Earth 0.7: One Simulates a Place, the Other Rebuilds a Real One

Put the two requests side by side and the difference is immediate. Ask Genie 4 for a rainy street at dusk and you get a street that has never existed, rendered as pixels, steerable by text, holding together for about two minutes. Ask ABot-Earth 0.7 for the same street and it will ask which street — or take a satellite image of one — then reconstruct that specific location as explorable 3D geometry in about ten minutes on a single consumer graphics card. They are both called world models and they are answering different questions: what could a place look like, versus what does this place look like.

That distinction decides almost every practical comparison between them. Genie 4, announced by Google DeepMind on October 27, 2026, is a generative simulator for scenes that do not exist. ABot-Earth 0.7, released by Amap — Alibaba's mapping and navigation division — on September 10, 2026, is a reconstruction engine for cities that do. Neither makes the other redundant, and choosing wrongly will cost you months.

The Two Requests, Stated Plainly

Genie 4 takes a prompt or an initial frame plus keyboard and camera input, and generates a video-like world in real time. You steer it with language mid-simulation: change the weather, move the camera, introduce a character. It generates audio natively — ambient sound, music, and dialogue with characters that respond in the scene — and it stays coherent for roughly two minutes before details drift. It is a research preview for trusted testers, with Google AI Ultra subscribers and Gemini API early-access partners next.

ABot-Earth 0.7 takes a georeferenced satellite image or a text description of a location and produces a 3D Gaussian splatting scene of that city — streets, buildings, terrain — which you can then explore. Amap positions it as the first 3D-native urban world model, capable of continuous generation from planet scale down to street level, across 196 countries and regions. It ships inside Amap's own consumer products: Flying Street View 2.0, Navigation Live, and a navigation feature the company calls the Avoid-Pitfall Guide.

Read those two paragraphs again and notice what is missing from each. Genie's output has no geography, no coordinates, no guarantee that the building on the left is the building that is actually there. ABot-Earth's output has no narrative, no characters, no sound, and no ability to become a place that is not already a place.

Pixels You Watch, Geometry You Walk

The output format is the deepest technical difference and the one most likely to determine your answer.

• Output — Genie 4 produces generated video frames with synchronised generated audio; ABot-Earth 0.7 produces 3D Gaussian splats, an explicit scene representation.

• Steerability — Genie 4 is driven by natural language at runtime; ABot-Earth 0.7 is driven by the choice of location at generation time.

• Geographic truth — Genie 4 has none by design; ABot-Earth 0.7 is anchored to a real coordinate.

• Audio — native and generative on Genie 4; not a feature of ABot-Earth 0.7.

• Time to a usable scene — seconds of generation on Genie 4, roughly ten minutes per square kilometre on ABot-Earth 0.7.

• Coherence limit — about two minutes of drift-free simulation on Genie 4; a static reconstruction that does not expire.

Gaussian splats matter here for a reason that has nothing to do with visual quality. A splat is a scene you can move a camera through without the model having to invent the parts you did not look at — because they were captured, not imagined. Genie 4 has the opposite property: everything off-camera is generated on demand, and Google concedes that object permanence drifts, which means the room you walked out of may not be the room you walk back into.

The Scale Claim, Translated

Amap's numbers are large and vendor-reported, and they should be read as such. The company says a square kilometre of city takes about ten minutes on one consumer-grade GPU, that this is up to 1,000 times faster than traditional photogrammetric reconstruction, and that it costs roughly one hundredth as much. Covering 196 countries is a coverage claim about inputs, not a claim that every one of those places has been reconstructed.

The 1,000× figure is the one to treat carefully. Traditional reconstruction pipelines — drone flights, photogrammetry, manual cleanup — take days and a survey crew, so a comparison against them is comparing a generation step to an entire field operation. Against other neural reconstruction methods, the gap will be far smaller. What is defensible is the shape of the claim: generating a kilometre of city in ten minutes on hardware you might already own is a different order of access from hiring a survey team.

Genie 4 publishes no comparable number at all. There is no resolution, no frame rate, no latency figure, and no stated cost. You cannot put its throughput next to Amap's, which is itself a finding: a model that will not state its frame rate is not yet competing on production terms.

Availability Is Where the Comparison Gets Honest

Both of these models are gated, and in different ways that matter.

Genie 4 is a research preview for trusted testers, with a public waitlist and a promise of wider access "in the coming weeks" through AI Ultra and the Gemini API. There is no price, no rate limit, no SLA, and no terms of use for commercial output. If you are a studio, you cannot plan around it; you can only watch it.

ABot-Earth 0.7 is worse for an outside developer despite being older. There are no public weights, no model card, no API, no pricing and no developer documentation. The experience site is Simplified Chinese only and generation is invite-only. It is a product feature inside Amap's consumer apps that the company has chosen to describe as a world model, not a platform you can build on.

So the honest answer to "which can I use" is: neither, today, without a relationship with the vendor. What differs is the direction of travel. Google has named the channels — AI Ultra, the Gemini API, AI Studio — and given a timeline in weeks. Amap has named its own apps.

Where Each One Is Actually Good

Genie 4's value is in worlds that must not be real. Film pre-visualization, where the point is to test camera language before a set exists. Game prototyping, where you want to explore a mechanic in a space no artist has built. Robotics and agent training, where Google's own demo paired Genie 4 with Gemini 4 Argon to decompose tasks for an Apptronik Apollo humanoid — the simulation is a rehearsal space, and a rehearsal space is more useful the less it resembles a specific street.

ABot-Earth 0.7's value is in worlds that must be real. Navigation products, where a wrong building is a wrong turn. Urban planning, where the question is what a proposed development does to a sightline. Insurance and inspection, where the deliverable has to correspond to an address. Anywhere the sentence "that isn't what it looks like" is a failure rather than a creative choice.

There is an obvious hybrid that neither company has shipped: use a reconstruction to establish geometry and geography, then use a generative model to restyle it, change its weather, populate it, or move a camera through it in ways the capture never covered. Amap has the geometry and no generative layer. Google has the generative layer and no geography. Whoever connects those two halves owns the next version of this category.

What It Would Take to Choose

If your work needs a specific real location, ABot-Earth 0.7 is the only one of the two that can give it to you, and you should start the conversation with Amap now rather than waiting for a developer programme that may never open.

If your work needs a plausible place that does not exist, with sound and live direction, Genie 4 is the stronger capability and the weaker product — a waitlist and a promise. The reasonable move is to prototype the surrounding pipeline against something you can call today and swap the world model in when access arrives. In our case that means building the planner and verifier half of the loop on routable LLM endpoints — one API, 0% markup on provider list pricing, automatic failover between providers — because that half is the part that will not change when the world model does.

The temptation is to ask which model is better. The useful question is narrower: does your world have to be somewhere? Genie 4 says no and gives you a camera. ABot-Earth 0.7 says yes and gives you coordinates. Almost every other difference between them follows from that single answer.