Skyfall AI unveils SKY-001, an engineering world model that turns text or sketches into executable CAD build steps, claiming 51x faster and 150x cheaper than GPT-6 Astra on CAD

Background and the industry pain point

Large models are strong at code and retrieval, but designing a physical product with hundreds or thousands of parts needs geometric and spatial reasoning plus very long-horizon planning. Most 3D generators output good-looking surface meshes with no part hierarchy, constraints or parametric feature tree, so engineers still decompose and model by hand. Skyfall AI, founded by the team behind Maluuba (the deep learning lab Microsoft acquired), argues this is a sequential decision problem rather than a language problem, and on October 19 released a research preview of SKY-001.

Architecture and mechanism

SKY-001 is a hierarchical visual world model and does not do next-token prediction. It works in four stages: generate several rough 3D options in parallel from text or a single 2D image; recursively decompose the chosen design with a diffusion-based process down to screws and fasteners; recreate each part the way a designer would, sketching 2D strokes and then extruding or cutting them into 3D; and assemble the parts hierarchically into a full assembly viewable in standard CAD tools such as Autodesk Fusion. In the Jaguar E-Type demo the car splits into 103 leaf parts and 139 nodes at depth 5.

Benchmarks

All figures are Skyfall's own and have not been independently reproduced. On 3D CAD, SKY-001 scores 79.7% IoU at 3.9 s and $0.005 per case; GPT-6 Astra 79.1% at 199 s and $0.75; Claude Opus 5.5 69.4%; Gemini 3.8 Flash 54.7%; Kimi K3 9.5%; Grok 4.7 completed none. On BIM, SKY-001 scores 43.6% at 1.22 s; the best LLM is Gemini 3.8 Flash at 42.3%, with GPT-6 Astra at 40.6% and 144.4 s. The "51-125x faster, 150-210x cheaper" claim in the launch post maps to the CAD and BIM charts on the blog. Eval set size, composition and parameter count are not disclosed.

Getting started

There is no API or open weights yet, only a waitlist on the website, and the press release targets commercial deployment in early 2027. Teams building CAD/BIM, hardware or digital-twin tools can study the 3D decomposition demos on the blog to judge whether a parts tree plus build steps fits their pipeline. Skyfall notes the final material finishes in the demo shots are applied afterwards, not predicted by the model, and the multipliers should not drive tool choices until you can test on your own parts.