Hi3D V1.0 vs ABot-Earth 0.7: One Sharpens an Object, One Generates a City

Ask what comes out the other end and the confusion disappears. Hi3D V1.0 returns a single mesh: one object, recentred to unit radius, geometry only, no texture. ABot-Earth 0.7 returns a kilometre of city as a 3D Gaussian-splatting scene, with roads, blocks and landmarks held in a consistent spatial frame. Both were released by Chinese vendors within the last six weeks. Both get filed under "3D model" in the same news cycle. Neither is a substitute for the other, and the useful comparison is not which one is better — it is which deliverable your pipeline can actually consume.

Two artifacts, not two competitors

Hi3D V1.0 is a refinement model from Hi3D, released on 23 October 2026 as the open-source successor to Sparc3D. It takes a reference image and a coarse mesh of the same object and adds geometric detail, in up to three passes, at requested resolutions of 512³, 1024³ or 1536³. The code and weights are downloadable; the licence permits commercial use with a scale trigger; the measured configuration is a 96 GB H20; and there is no hosted inference endpoint.

ABot-Earth 0.7 is a 3D-native urban world model from Amap, Alibaba's mapping and navigation arm, unveiled on 10 September 2026 at its Street Stars event in Hangzhou. It accepts a georeferenced satellite image or a text prompt and generates an explorable city scene in 3D Gaussian Splatting format. Amap says a kilometre-scale scene takes about ten minutes on one consumer-grade GPU, that coverage reaches 196-plus countries and regions, and that the output exports to Unreal Engine and Unity. Those are vendor figures. No weights, model card, public API or pricing have been published, and generation on the experience site is described as invite-only.

Put those two paragraphs side by side and the comparison is already doing something more useful than a spec sheet: it is separating a tool from a product demonstration.

The input is the first asymmetry

ABot-Earth 0.7 is end-to-end on the input side. One satellite image or one text prompt goes in; a scene comes out. You do not supply geometry, and Amap does not ask you to.

Hi3D V1.0 is a second stage. It needs a coarse mesh before it can do anything. The repository ships an optional demo adapter that generates one with Tencent's Hunyuan3D-2 mini, but the core refinement script does not import Hunyuan and will accept any coarse mesh you already have. That design is deliberate — it lets the model slot behind whatever generator you use — but it means the honest description of Hi3D is "an upscaler for geometry," not "an image-to-3D model," even though the Hugging Face tags include image-to-3D.

This is the first place a buyer should stop. If your workflow is "one image in, asset out," Hi3D alone does not give you that. If your workflow already has a coarse mesh and a reference render — which most asset pipelines do — Hi3D is aimed exactly at the gap between the two.

Six dimensions that decide the purchase

• Unit of output — Hi3D V1.0 one object vs ABot-Earth 0.7 one kilometre-scale urban scene

• Format — Hi3D V1.0 a geometry-only trimesh mesh, exportable as GLB vs ABot-Earth 0.7 3D Gaussian Splatting, with Unreal Engine and Unity export claimed

• Surface appearance — Hi3D V1.0 no texture, material or vertex colour vs ABot-Earth 0.7 textured scene output, per Amap

• Input — Hi3D V1.0 a reference image plus a coarse mesh vs ABot-Earth 0.7 a satellite image or a text prompt

• Compute — Hi3D V1.0 measured on a 96 GB NVIDIA H20, no CPU path vs ABot-Earth 0.7 about ten minutes per square kilometre on one consumer-grade GPU, vendor-reported

• Access — Hi3D V1.0 code and weights downloadable today under a custom commercial licence vs ABot-Earth 0.7 no weights, no API, no pricing, invite-only beta

A splat is not a mesh

This is the part that most comparisons miss, and it is not a technicality. A Gaussian-splat scene is a collection of oriented, coloured blobs that render beautifully and carry no explicit surface. A mesh is a connected set of triangles with a boundary. The two are good at different things.

Splats render. They can be viewed from novel angles, streamed at level of detail and dropped into a real-time engine. They are a poor input to a slicer, a CAD kernel, a collision mesh or a rigging pipeline, because those tools want a surface and a watertight volume. ABot-Earth's output is on the render side of that line. Amap's claim of Unreal and Unity compatibility does not specify whether the export is a splat asset, a mesh or both, and no public format documentation exists to check. That is not a flaw in the model; it is a question a buyer must ask before building a pipeline around it.

Hi3D's output is on the surface side. A trimesh mesh can be exported, sliced, booleaned and measured. But the repository normalises it to unit maximum radius, discards world coordinates and adds no texture. For 3D printing, that is a usable starting point after a rescale. For e-commerce or a game asset, it is half a deliverable: the other half is a texturing step that Hi3D does not perform.

So the format row is not "mesh versus splat, and mesh wins." It is "the artifact you can post-process versus the artifact you can look at." Choose the one your next tool accepts.

Compute and access, which is where the comparison gets lopsided

Hi3D is expensive per object and cheap per experiment. The measured configuration is a 96 GB H20, and the README says the full 1536³ ladder has only been validated there. Batch size is one. There is no CPU path, and flash-attn has no fallback. A 512³ request runs one pass at 25 scheduler updates; 1536³ runs three. The repository publishes no wall-clock timing, so you cannot budget a production line from the card. What you can do is download it today, run it on hardware you already have, and measure.

ABot-Earth 0.7 is cheap per scene on paper and unavailable in practice. Ten minutes per square kilometre on a consumer GPU is a striking claim — Amap also says up to 1,000× faster and about one hundredth the cost of traditional reconstruction — but every one of those numbers comes from Amap and none has been independently reproduced. More to the point, there is no way for an outside team to reproduce them: no weights, no model card, no API, no pricing, no developer documentation. The experience site is live, the interface is Simplified Chinese only, and generation is whitelisted. You are comparing a downloadable model against a press cycle.

That asymmetry decides the quarter even if ABot's technical claims are entirely true. A team that needs urban 3D this month cannot procure ABot-Earth 0.7. A team that needs object geometry this month can install Hi3D V1.0 and find out what it does.

What a team would actually do with both

There is a real composition, and it is worth stating precisely because it is easy to overstate. A city scene needs an environment and it needs props. ABot-Earth 0.7, if it were available, would supply the environment: terrain, blocks, building massing, street layout. Hi3D V1.0 supplies a better version of an individual prop, given a coarse mesh and a reference image. Assemble the two in an engine and you have a scene with detailed assets inside a generated city.

The seam is not free. ABot's scene is a splat representation with its own coordinate frame; Hi3D's mesh is normalised to unit radius and has no absolute scale. Putting a refined prop at the right place, at the right size, in a splat scene requires a transform and a scale convention that neither vendor publishes. Nobody has written that integration up. It is the kind of work a studio does in a week and keeps to itself.

For most teams the more realistic split is simpler: use ABot-Earth 0.7 as evidence of what Amap's spatiotemporal data can do, not as a dependency, and use Hi3D V1.0 where you already own the coarse-mesh stage.

The bill neither page shows

Neither model is served by OrcaRouter. The catalogue covers text, vision, video, embedding and speech models, not 3D geometry, and this article is not going to pretend otherwise. What both pipelines do consume is language and vision calls: expanding a text prompt into a usable scene description, captioning reference renders, checking whether an output matches the brief, writing the metadata that makes an asset findable.

That is the layer OrcaRouter covers. More than 200 models sit behind one API, the provider's list price is passed through at 0% markup, and automatic failover keeps a degraded provider from stopping a batch halfway through. For a studio running hundreds of asset jobs, the difference between one key and five is not elegance; it is how many places a job can fail.

The decision, without the hedging

If the thing you need to hand someone is a printable or editable object, Hi3D V1.0 is the only one of these two you can install, and the only one with a licence that contemplates commercial use. Budget for the upstream coarse mesh, the 96 GB card and a texturing step. If the thing you need is a walkable city, ABot-Earth 0.7 is the more interesting system and the less available one; treat it as a benchmark for what a well-resourced mapping company can generate, not as a component you can schedule.

The comparison only looks strange if you assume "3D model" is one category. It is at least two: geometry you can manufacture, and environments you can inhabit. Hi3D and ABot-Earth sit on opposite sides of that line, which is why the honest answer to "which is better" is another question — what does your next tool accept?