By Hamza Faisal

World models have spent the last two years learning to make video games, synthetic footage, and robot training scenes. Skyfall AI thinks they have been practicing for the wrong job. On Monday, the startup's co-founder and CEO Sam Pasupalak — the former chief executive of Maluuba, the deep-learning lab acquired by Microsoft Research — unveiled SKY-001, which he calls the first "Engineering World Model."

Instead of predicting the next frame of a video, SKY-001 predicts the next step of a build. Feed it an image of an engine, a line of text, or a building outline, and it unfolds the object into a hierarchy of assemblies and parts, then produces the ordered visual sequence — the 2D sketch and the machine operation — that turns raw material into each finished piece. Skyfall says 18 months of work, built by a team out of Maluuba's deep-learning lab, went into the model.

One photo in, a parts tree out

The demo the company is leading with is deliberately hard: a single photo of a VMS-85 engine. SKY-001's Hierarchical Composer produced a five-level part tree from it — down to sleeves and shafts hidden inside housings that are not even visible in the input image. A second model, the Construction World Model, then rolled out each leaf part as ordered visual states: the solid as it looks after each operation, plus the 2D sketch that turns one state into the next.

That two-step structure is the point. General-purpose chatbots can describe a CAD part; SKY-001 is built to plan one, part by part, operation by operation. Where building-information-modeling and CAD workflows demand precise geometry, the model hands engineers a build sequence they can inspect rather than a finished render they have to trust.

150x cheaper — on Skyfall's own scoreboard

The headline numbers are aggressive. Skyfall claims that on CAD and BIM tasks, SKY-001 is 51 to 125 times faster and 150 to 210 times cheaper than the frontier models it tested against — including GPT-6 Astra — while matching or exceeding their accuracy, measured by IoU (a standard overlap score for predicted shapes). Independent reporting on the announcement puts the design-cost advantage for CAD and floor-plan work at 150x to 185x.

The caveat is the usual one: these are the company's own benchmarks, released via an X thread, and the detailed test conditions remain limited. Independent validation and performance inside real engineering workflows — where a wrong tolerance is worse than a slow one — will decide whether SKY-001 can progress from visual planning to actual manufacturing use. For now, access is being offered through a waitlist.

The quiet race for the AI general engineer

What makes this launch interesting is not the size of the model but the target. CAD and BIM are verifiable tasks: a part either fits or it doesn't, and a build sequence either produces the object or it fails. That makes engineering a natural arena for world models, which have historically been judged on the far softer metric of whether generated video "looks right."

Pasupalak's pitch is that the first "AI General Engineer" will be the backbone of autonomous manufacturing — letting engineers dream a physical object before they build it. If even a fraction of the claimed cost advantage holds up under third-party testing, engineering software is about to get its diffusion-model moment.

By Hamza Faisal