Hi3D v1.0 vs SKY-001: One You Can Run, One You Cannot

Hi3D v1.0 and SKY-001 both claim to serve engineering and manufacturing, and they attack the same job from opposite ends. SKY-001, Skyfall AI's "engineering world model" announced on 19 October 2026, takes text or a 2D image and proposes rough 3D assets plus a staged CAD or BIM assembly sequence. It wants to make the first vertex. Hi3D v1.0, open-sourced on 23 October 2026, cannot make a first vertex at all: it requires a coarse mesh and a reference image, and returns that mesh with more detail. One proposes structure; the other finishes it. The practical difference is that you can download one of them today.

That availability gap is not a footnote. It changes what the comparison is for. With Hi3D v1.0 you can run your own test this afternoon and decide from evidence. With SKY-001 you are deciding from a press release, a vendor benchmark and a waitlist.

The first-vertex problem

Every 3D asset pipeline has a bootstrap step where something decides the topology, the proportions and the part count. It is the most consequential decision in the pipeline and the hardest to automate, because there is no single correct answer: a bracket can be one part or six, and both can be right.

SKY-001 is an attempt to automate that decision from a description. Hi3D v1.0 explicitly declines it. The README's limitation section says "coarse structure matters," that severe coarse-mesh errors can persist, and that hidden geometry is only plausibly completed. In other words, Hi3D will make your structure more detailed but will not tell you the structure is wrong. That is a reasonable division of labour. It also means the two models are sequential, not interchangeable, and any comparison that scores them on the same axis is measuring the wrong thing.

What SKY-001 says it produces

The claimed pipeline is text or a 2D image in, several rough 3D assets out, followed by a staged construction plan. A hierarchical composer decomposes a design into systems, assemblies, subassemblies and parts — or, on the building side, floors, units and rooms — and a construction world model plans ordered visual states rather than emitting a single final mesh. Sketch generation runs through a diffusion transformer in two stages, and a Matryoshka autoencoder compresses the geometric representation so that planning is cheaper.

That is a design tool, not a mesh generator in the Hi3D sense. The output that gets scored is a proposed assembly evaluated by volumetric overlap against reference geometry. Whether it is watertight, manifold, tolerance-aware or directly manufacturable is not what the benchmark measures, and Skyfall has not released weights, a licence, an API or a public test set to check.

What Hi3D v1.0 actually returns

A geometry-only trimesh.Trimesh, recentered and normalized to unit maximum radius, with no textures, materials or vertex colours. You choose one of three settings: 512³ runs one pass, 1024³ runs two, 1536³ runs three, at 25 scheduler updates per pass by default. The validated machine is a 96 GB NVIDIA H20; there is no CPU path, batch size is one, and the 1536³ tier has not been validated anywhere else.

Crucially, Hi3D has no opinion about what the object is supposed to be. It knows what the photograph shows and what the coarse mesh asserts. If the coarse mesh has a bracket where the design needs a boss, Hi3D will refine the bracket beautifully. The model is a finisher, and finishers inherit the brief.

The benchmarks, with the caveats attached

• SKY-001, CAD track — 79.7% IoU at 3.9 seconds and $0.005 per case, against GPT-6 Astra at 79.1% IoU, 199 seconds and $0.75. Vendor-reported, on an internal test set of about 1,000 components and roughly 100 steps each, with no public split.

• SKY-001, BIM track — 43.6% IoU against 40.6%, at 1.22 seconds against 144.6 seconds. Same provenance, same caveat.

• Hi3D v1.0, slicing — 90% slicing success, stated in the press release and absent from the repository. No slicer is named, no failure definition is given, no baseline is reported, and no test meshes ship with the code.

• Hi3D v1.0, fidelity — "highest multi-view fidelity and detail richness in the reported evaluation." The evaluation is not in the repo, and there is no comparison table or competitor list.

Neither set of numbers has been independently reproduced. That is the honest common ground, and it is worth saying plainly rather than picking the side whose table looks more precise. The difference in kind is what the claim is attached to: Hi3D's 90% sits next to a runnable artifact, so you can design a test that falsifies it. SKY-001's 79.7% sits next to a waitlist, so you cannot.

The narrower point is that a 0.6-point IoU lead over GPT-6 Astra is not a reason to move a pipeline. Without a published test set it is within the range that a different sampling of components could erase. The 51× to 125× speed claim is more interesting, because it is large enough to survive some measurement error, but it is still a vendor measurement on vendor hardware.

What IoU does not tell a manufacturer

Volumetric overlap measures how much of the predicted shape coincides with the reference shape. It does not measure wall thickness, minimum feature size, tolerance stack-up, surface finish, draft angle, or whether an assembly sequence can physically be executed. A part can score well on IoU and still be unmanufacturable, and a part can score poorly and still be printable with a small edit.

The same translation problem applies to Hi3D's slicing figure. Slicing success is not print success. A file can slice cleanly and fail on the bed, and a file that slices with repairs can print perfectly well. Neither vendor has published the downstream test that would connect its headline number to a physical result, which is why both numbers should be read as claims about software behaviour rather than manufacturing readiness.

Decoding the cost claims

• SKY-001's $0.005 per case — an internal benchmark-run compute estimate, not a price. Skyfall has published no API, no pricing and no commercial terms. Comparing it to GPT-6 Astra's $0.75 is comparing an estimated internal cost to a list price, which flatters the cheaper side twice: once for the estimate and once for the absence of a margin.

• Hi3D v1.0's cost — the weights are free under a custom licence, and the real cost is a 96 GB GPU and the minutes it takes to run three passes. No latency table has been published at any resolution, so the per-asset cost is something you measure rather than look up.

The honest comparison is not $0.005 against $0.75. It is "a waitlist with no price" against "a GPU bill you can estimate today." Those are different kinds of unknown.

A pipeline that would use both, if SKY-001 ships

The natural composition is easy to describe: SKY-001 proposes the assembly and part breakdown, the rough geometry is exported, Hi3D v1.0 refines each part to a printable density, a repair step fixes topology, and the result goes to manufacturing. That is a coherent division of labour, and it is the reason the two models are worth comparing at all.

It is also entirely aspirational today. SKY-001 has no export path, no weights, no API and no licence, and Skyfall is targeting commercial deployment in early 2027. Hi3D v1.0 has the opposite problem: it is available now but needs an upstream stage that SKY-001 has not shipped. Anyone building the pipeline this quarter is building it around a waitlist.

OrcaRouter is not part of either 3D stage and does not host these models. It is useful for the text and vision work that surrounds them — parsing a specification into a structured brief, reading a drawing, or running a panel of vision models over a render to catch defects before a human reviews it. That last case is where model fusion earns its keep: several models answer the same inspection question and you look at where they disagree. The same endpoint handles it with automatic failover, so an inspection step does not become the fragile part of an otherwise local pipeline.

What would settle it

For SKY-001, three things: a public test set, weights or at least an API, and a benchmark that measures manufacturability rather than overlap. Until then the 79.7% is a press-release number with a credible architecture behind it and no way to check the arithmetic.

For Hi3D v1.0, the missing piece is smaller and more specific: a named slicer, a published mesh set, a failure taxonomy, and the unrefined coarse mesh as a baseline so a reader can see what the refinement added. That is a weekend of work for the vendor and it would convert the strongest claim in the release into something a print shop could act on.

Until either lands, the decision rule is unusually simple. If you need geometry today and you have the GPU, Hi3D v1.0 is the only one of the two you can actually test. If you need someone to propose the structure in the first place, SKY-001 is the more ambitious idea and, for now, the less usable product.