Sandflow's sand3d Lets AI Generate 3D Objects Using 6x Fewer Tokens

AI systems that emit raw mesh data spend tokens describing thousands of vertices while discarding the object’s editable structure. Sandflow has released sand3d on GitHub , an Apache 2.0 library that represents a 3D object as a short Python program with named parts, edges, faces and joints. A model can generate the program, a developer can inspect it, and either can revise individual components later.

Built on Blender’s bpy module, sand3d exposes a flat set of functions for drawing profiles, creating solids, editing geometry, positioning frames, defining joints and assembling parts. Those functions form a compact modeling language while the resulting object remains an ordinary Blender scene.

Python becomes the 3D format

Sandflow’s benchmark reports that sand3d uses between one-half and one-sixth as many tokens per part as equivalent Blender Python, CadQuery or FeatureScript code. The complete API reference occupies about 3,600 model tokens, making it small enough to include in a system prompt alongside instructions and the requested object specification.

The API builds geometry through operations that map cleanly to modeling concepts. A profile defines a 2D outline, an extrusion turns it into a solid, and operations such as holes and fillets modify named regions. Developers and models work with labels such as plate and top , reducing the coordinate bookkeeping common in procedural 3D scripts.