Step-by-step room design over single-block scripting: SKY-001 applies World Models to CAD

Designed by Maluuba alumni, SKY-001 plans an object piece by piece and matches GPT-6 Astra in CAD according to Skyfall, at 150 times lower cost

Language models reason in text, a medium that startup Skyfall AI deems ill-suited to maintaining consistency across the thousands of components in a physical object. Based in San Francisco with a research hub in Toronto, the company is introducing SKY-001 as the first World Model dedicated to engineering. Its founders—Sam Pasupalak, Sumit Pasupalak, and Kaheer Suleman—come from the team at Maluuba, a deep learning lab acquired by Microsoft, and its investors include Fidelity, Touring Capital, Inovia, and M13. The model does not generate an environment to explore: according to the announcement post, it produces the design of an object and the sequence of steps to build it, in CAD for industrial products as well as BIM for floor plans. Breaking Down the Object Before Drawing It

The system chains two modules together. The first, named the Hierarchical Composer, starts from an image of the product, generates a rough 3D shape, and then breaks it down into assemblies and each assembly into parts, down to single individual components, which a final pass completes and details. Taking a photographed engine as an example, the resulting tree has five levels, ranging from the block and cylinder heads down to the gears and sleeves hidden inside. The second module, the Construction World Model, predicts for each part the sequence of intermediate 3D states that bring it to life—an extrusion, an addition of material, a cut—and draws the 2D sketch between two states that transitions from one to the other.

This hierarchy explains the claimed speed advantages. Once a parent is established, its elements are constructed in parallel, so that the number of successive steps depends on the depth of the tree rather than the number of parts. Sketches are first generated by diffusion in a highly compressed latent space that preserves only the general layout, and then a refinement module redraws them in full resolution. In BIM, the same approach transforms the outline of a floor and a spatial program into apartments, then into rooms, and drafts a complete floor plan in about one second. 288 Parts per Object at the Median

To train the Hierarchical Composer, Skyfall cleaned public datasets of real product assemblies, whose trees were inconsistent and whose parts were often duplicated or unnamed, and then reconstructed them according to industrial design practices. The result averages a median of 288 parts per object, compared to 6 for PartVerse—a common benchmark for part decomposition—with over 200,000 parts for the first module and over 600,000 construction examples for the second. When a part's construction history was missing, engineering tools and LLM-based pipelines reconstituted it prior to verification. Meanwhile, the BIM models learn from approximately 35,000 real and synthetic plans, representing some two million drawing states. Benchmark Scores Measured by Skyfall

On a CAD benchmark of 1,000 components drawn from complex objects—a Jaguar E-Type, a FANUC robotic arm, a Bell 222 helicopter, or a jet engine—each case requires about a hundred drawing steps, and the result is scored by IoU, the overlap ratio between the generated shape and the reference. The model achieves 79.7% in 3.9 seconds and at $0.005 per case, compared to 79.1% in 199 seconds and $0.75 for GPT-6 Astra, representing roughly 50 times faster execution and 150 times lower cost at virtually identical accuracy. Behind them, Gemini 4 scores 76.9%, while the other tested models fall off significantly and Grok 4.7 fails to produce any result.

In BIM, across 1,000 floor plans, the speed gap compared to GPT-6 Astra widens to about 120 times and the cost gap to about 185 times, with an IoU of 43.6% versus 40.6%, though Gemini 3.8 Flash follows closely at 42.3%. The launch post published on 𝕏 mentions speed gains of 51 to 125 times and cost gains of 150 to 210 times—upper bounds that the numbers in the blog post do not replicate.

For part decomposition, two judge models, GPT-6.1 Sol and Claude Opus 5.5, blindly evaluated the renders of SKY-001 and GPT-6 Astra, the latter writing a Blender script for the entire assembly in a single prompt. Scores out of 10 favor SKY-001 across all five criteria, ranging from 7.8 vs. 6.4 for decomposition to 7.7 vs. 7.4 for exterior realism, with GPT-6 Astra rendering a model for only 82% of the objects. All of these benchmarks were designed and run by Skyfall, and the post makes no mention of publishing the test sets. A Visual Plan, Not Yet an Editable File

In its current form, SKY-001 delivers a visual design plan: a part tree and sequences of images showing how each part is built. The blog post notes that its transition animations are illustrative and that the saved sketch may diverge from the 3D geometry. Converting these sketches into executable operations, complete with an editable feature history, is listed among the announced next steps, as is converting floor plans into walls, doors, and windows within a BIM model that architects could edit. On 𝕏, Sam Pasupalak also mentions pairing the model with a vision-language-action agent tasked with executing each step, verifying the outcome, and correcting whatever fails. The model remains in research preview, available via a waitlist, with no API, pricing, or model weights released, and Skyfall is targeting a commercial rollout in early 2027 according to The AI Insider.