Hy-Nomas vs SKY-001: One Renders a World, the Other Assembles It

Tencent published 280 benchmark cases for Hy-Nomas and no scores. Skyfall published scores for SKY-001 and no released test set. The symmetry is neat enough to look like a coincidence, and it is the most useful thing about putting these two models side by side, because beyond the phrase "world model" they have almost nothing in common. Hy-Nomas, released by Tencent Hunyuan between 21 and 23 October 2026, generates explorable video: 720 × 1280 at 24 FPS, steered by free-form text and camera trajectories. SKY-001, announced by Skyfall AI on 19 October 2026, generates engineering geometry: it takes a text prompt or a 2D image and produces an ordered CAD or BIM construction sequence. One model's output is something you look at. The other's is something you manufacture.

The distinction is not pedantry. It decides what accuracy means, what a benchmark can measure, whether you can obtain the model at all, and what you would actually do with it on a Monday morning.

Two jobs behind one phrase

A video world model predicts the next observable state of a scene. Its unit of output is a frame or a segment, and its correctness criterion is whether the result looks and behaves like a world. An engineering world model predicts the next structural state of a design. Its unit of output is a part, an assembly or a construction step, and its correctness criterion is whether the geometry matches what was asked for and whether the sequence can be built.

Hy-Nomas is segment-autoregressive. It renders a span of video, keeps it in a lifelong KV cache, takes the next instruction, and renders the next span. Text and camera are separate inputs that compose: the prompt describes the scene, the extrinsics describe where the lens goes. SKY-001 is hierarchical. A composer decomposes the request into systems, assemblies, subassemblies and parts, then down through floors, units and rooms; a construction world model plans the ordered visual states that produce each element. Both call themselves world models because both learn a dynamics model of a domain. The domains do not overlap.

• Output — Hy-Nomas video segments at 720 × 1280 and 24 FPS; SKY-001 structured CAD/BIM construction sequences targeting design tools

• Input — Hy-Nomas free-form text plus camera extrinsics; SKY-001 a text prompt or a 2D image

• Definition of correct — Hy-Nomas perceptual coherence across turns; SKY-001 geometric overlap against a reference design

• Unit of evaluation — Hy-Nomas a session; SKY-001 a case

• What actually ships today — Hy-Nomas Apache 2.0 weights, about 196 GB; SKY-001 a waitlist place

What SKY-001 claims, and what the claim is worth

Skyfall's announcement is a vendor benchmark, and it is a well-constructed one: two tracks, a thousand cases each, a named comparator, and enough detail to see the shape of the test. The CAD track runs 1,000 components at roughly 100 drawing steps per case. The BIM track runs 1,000 floor plans at an average of about 70 steps. Against GPT-6 Astra, Skyfall reports 79.7% versus 79.1% mean IoU on CAD at 3.9 seconds versus 199 seconds and $0.005 versus $0.75 per case. On BIM it reports 43.6% versus 40.6%, at 1.22 seconds versus 144.6 seconds and roughly $0.0024 versus $0.446.

Three things follow from reading the numbers rather than the headline. First, the "150x cheaper" figure is a benchmark-run cost ratio, Skyfall's estimate of what it cost to execute one case, not a price anyone can buy at. There is no public SKY-001 price list. Second, IoU measures geometric overlap, not engineering quality: it does not tell you whether a part can be manufactured, whether tolerances close, or whether a floor plan satisfies egress, accessibility or structural requirements. Third, the reported timings exclude upload, preprocessing, queueing, CAD or BIM conversion, validation, export and human review, and every one of those sits in a real deployment.

The accuracy gap is the part that gets lost. On CAD it is 0.6 percentage points across a thousand cases; on BIM it is 3.0. That is a real result and a narrow one. It establishes that a small team's specialised model can land in the same range as a frontier general model on this task, not that it replaces one. Skyfall's own framing is closer to the truth than the press coverage: the company describes SKY-001 as a research preview with access through a waitlist and commercial deployment targeted for early 2027, not as a product you can call.

What Hy-Nomas ships, and what it refuses to claim

Tencent's release is the opposite artifact: a complete kit with the conclusion removed. The weights are Apache 2.0 and they include more than a checkpoint. There is a 54-block dual-stream MMDiT at hidden size 2,048 with 16 attention heads, a Qwen2.5-VL-7B-Instruct text encoder, 32-channel DCAE-V latents, a fine-tuned PixelUMM decoder in FP32 and FP8, and eight denoising steps in the distilled configuration, with four and six also accepted. Two LoRA adapters ship: Cinematic, and Robotics trained on the RealSource-World robotic-arm dataset. The repository documents how to train your own: rank and alpha 256, zero dropout, the Muon optimizer at a constant learning rate of 1e-4 for 2,000 steps, trained from the FP32 checkpoint and merged into the distilled one.

What does not ship is a number. Hy-Nomas-Bench has 280 cases across four tracks, multi-turn text control, camera navigation, interleaved text-and-camera control and spatial memory, with about 220 reference camera trajectories. Its results table is empty. The technical report link on the model card is marked coming soon and resolves to nothing. The Hugging Face page states that the model is not deployed by any inference provider. The GitHub repository was ten hours old, at 21 stars and zero releases, when this was written.

The asymmetry that decides most real cases

If both models were equally good at their own jobs, this would still not be a close call for most readers, because only one of them can be obtained. Hy-Nomas is a download: run it, fine-tune it, serve it. The caveat is that "run" means eight GPUs on the documented path, or a single RTX 4090 with INT8 DiT weights, an INT8 KV cache, an FP8 decoder, a sliding-window history cap and CPU offload. The full checkpoint set is about 196 GB; the low-memory route needs roughly 38 GB plus the text encoder. It is not a laptop download, and there is no hosted endpoint from Tencent to fall back on.

SKY-001 is a waitlist. No weights, no license terms, no public API, no published price, and a commercial target of early 2027 that the company has not committed to as a date. For a reader deciding what to do this quarter, that difference does more work than any benchmark on this page.

It also reframes the question. "Hy-Nomas versus SKY-001" is not a choice between two ways to do the same thing. It is a choice between two capabilities, one of which you can have today and one of which you cannot yet have at all.

Where the two vocabularies actually touch

Robotics is the one place they come close. Hy-Nomas's Robotics LoRA was trained on a robotic-arm dataset and merges into the distilled checkpoint to improve manipulation scenes. SKY-001's composer works on mechanical assemblies and parts, which is the same physical vocabulary from the other direction. A lab could plausibly design a bracket with an engineering model and simulate an arm picking it up with a video world model, and the combination is more interesting than either piece.

But the interfaces do not exist. SKY-001 emits construction sequences aimed at CAD tools. Hy-Nomas consumes text and camera trajectories. Nothing in either release speaks to the other, neither has been validated against physical hardware in a published test, and the robotics claim on both sides is a capability demonstration rather than a measured result. The meeting is a diagram, not a pipeline.

The cost conversation people get wrong

Two kinds of number are circulating about these models, and neither is a price. SKY-001's $0.005 per CAD case and $0.0024 per BIM plan are internal benchmark-run estimates, and the $0.75 and $0.446 comparators are the cost of running GPT-6 Astra through the same harness. They are useful for judging efficiency. They are not rates, and Skyfall has published no rate card.

Hy-Nomas has no cost figure at all, because there is no service to price. Its cost is GPU hours: one RTX 4090 with the low-memory configuration, or eight GPUs on the standard path, plus the download and the storage. If you want a truthful comparison of what each costs to operate, you need your own utilisation assumptions, and neither vendor has given you the numbers to make them.

The one thing that is comparable is what each costs to evaluate. Hy-Nomas lets you run its 280-case benchmark yourself, because the dataset is published; you will have to build the scoring harness, since Tencent did not ship results. SKY-001's benchmark is described but not released, so a third party cannot reproduce it without rebuilding the test. That asymmetry is worth more than the $0.005 headline.

The model is not the workflow

Neither model is on our catalogue, and neither is likely to be soon: one is a self-hosted open-weights release, the other is a waitlisted research preview with no API. We are not going to pretend otherwise.

What both of them need, and neither of them includes, is the reasoning layer around the generator. A SKY-001-style workflow starts with a request that has to be decomposed into a design tree, routed to the right tool, checked against constraints and written back into a CAD file. A Hy-Nomas session starts with a natural-language intention that has to become a camera path and a segment-level instruction. Those are language-model jobs, and they are where reliability is won or lost. OrcaRouter gives that layer one endpoint for more than 200 models, with 0% markup on provider list price, so a cheaper model can handle decomposition while a stronger one handles the final check, and automatic failover keeps the chain alive when a provider degrades. You can express the whole sequence in the routing DSL instead of hard-coding three integrations. The world model does the rendering or the assembly; the routing layer decides what to ask it for.

What would make this a real comparison

The comparison becomes meaningful the moment either side closes its gap. If Tencent publishes scores for Hy-Nomas-Bench, the model gets a baseline and every future world model can be measured against it. If Skyfall releases its CAD and BIM test sets, the 79.7% becomes reproducible rather than reported, and the 0.6-point lead gets the scrutiny a 0.6-point lead deserves.

Until then, the correct reading is categorical rather than competitive. Hy-Nomas is an open, self-hostable video world model with an empty results table. SKY-001 is a closed, waitlisted engineering world model with a vendor table and a narrow lead. They share a phrase, a month and a methodological gap. They do not share a job.