What if AI could generate an entire 3D model by writing a few hundred tokens of Python?
Not just a beautiful image. Not an uneditable collection of polygons.
An actual parametric 3D object with named components, adjustable dimensions, and reproducible geometry.
Meet sand3d by Sandflow AI, a new open-source Python library accompanied by an enormous dataset of 3D models designed specifically for training large language models.
And this is exactly the kind of development I find exciting about the future of AI-powered 3D.
The numbers are impressive:
106,094 3D objects converted into a unified code representation
10,456 articulated models with moving parts
Five source datasets, including Objaverse, CAD, Articraft, NVIDIA SimReady, and AutoMate
3.8× fewer tokens than conventional Blender Python in the team's CAD comparison
Fully reproducible geometry generated from short Python programs
Open-source library released under Apache 2.0
But here's the really interesting part.
These aren't ordinary 3D meshes. They're programs that describe how objects are built.
Traditionally, many publicly available 3D datasets contain enormous collections of polygons.
Thousands of triangles. Inconsistent topology. Objects that look convincing but are difficult to edit, animate, or understand structurally.
For AI models, learning from that kind of data can be challenging.
Sandflow took a different approach.
They reconstructed more than 100,000 existing 3D objects as compact Python programs.
And each program describes the object in meaningful parts.
Imagine a mechanical assembly.
Instead of receiving an anonymous mesh containing thousands of triangles, an AI model can work with:
Named components and their relationships
Adjustable dimensions and proportions
Geometric operations such as extrusion, cutting, and filleting
Materials and object placement
Mechanical joints, movement axes, and limits
Complete assemblies built from smaller parts
Want to change the width of a bracket?
Change a parameter.
Need a larger opening in a mechanical component?
Adjust the corresponding dimension.
Want to modify a moving mechanism?
Edit the joint definitions and regenerate the geometry.
That's the beauty of parametric 3D.
And Sandflow designed its language to be particularly efficient for LLMs.
In a comparison involving 30 CAD parts, the team reports that sand3d required 3.8 times fewer tokens than equivalent Blender Python scripts.
That's approximately 74% less code for the model to process.
The library also supports exporting geometry into formats including GLB, STL, OBJ, URDF, and OpenUSD.
And here's why this matters.
We've spent years teaching AI models to generate images, videos, textures, and 3D meshes.
But creating a visually convincing model isn't the same as creating something that can be reliably modified or used in a professional production pipeline.
Parametric code gives AI a much more structured way to understand and create geometry.
And with more than 100,000 examples now available for research and training, developers have a substantial dataset to experiment with.
Imagine asking Claude Code, Codex, or another LLM:
"Create a mechanical assembly with adjustable dimensions, movable joints, and separate components suitable for exporting to a CAD workflow."
Instead of producing an enormous Blender script filled with hard-coded coordinates, the agent could use compact, reusable geometric instructions.
Of course, this doesn't mean every LLM can suddenly generate flawless engineering-grade CAD models.
The converted objects still need quality checks, and the dataset's usefulness for training future models will need to be demonstrated experimentally.
But the direction is fascinating.
We're moving from AI that generates the appearance of objects toward AI that can describe how those objects are actually constructed.
And personally, I've been expecting this transition for quite a while.
Maybe the future of AI-generated 3D won't be another giant model producing increasingly beautiful triangle soup.
Maybe it'll be a language model that understands geometry well enough to build an editable object directly through code.
And with sand3d, we're getting another important piece of that puzzle.
Sand3D source code:
https://github.com/Sandflow-ai/sand3d
106,094-object dataset collection:
https://huggingface.co/collections/Sandflow/sand3d
Interactive 3D Explorer:
https://huggingface.co/spaces/Sandflow/sand3d-explorer
Official Sandflow project:
https://sandflow.ai/sand3d/
Technical dataset breakdown:
https://huggingface.co/blog/Sandflow/sand3d-106094-3d-objects-in-one-code-substrate-for
#Sand3D #SandflowAI #AI3D #ParametricModeling #AICAD #ClaudeCode #GenerativeAI #OpenSourceAI
YouTube / ABVX