sand3d lets LLMs write 3D models as short Python programs
A Blender-backed Python library turns parts, joints and assemblies into brief named programs, with one-call exports to GLB, URDF, USD and IFC.
sand3d, a library for 3D modeling in Python built for LLMs, had 1,969 stars and 218 forks on GitHub as of 2026-11-14, 11 days after the repository was created on 2026-11-03. It ships under Apache-2.0, and GitHub lists HTML as the main language. The premise is narrow and specific: shapes are written as short Python programs over Blender's bpy, so a model can produce a part in a few hundred tokens and a person can read what it did.
Parts as programs, not coordinates
Every verb sits on one flat s3d. surface. Profiles cover rect, circle, slot, polygon, arc, spline, superellipse and text among 16. Solids and paths add extrude, revolve, sweep, loft, helix, tube and torus among 15. Edits include cut, join, intersect, split, hole, pocket, fillet, chamfer and shell, again 15. Frames, articulation (joint, pose, seat), assembly verbs and part generators such as gear, knurl and thread fill out the rest.
Two design choices carry the claim. Parts, edges and faces are reached by name rather than by hand-computed coordinates, and the same program is said to build the same mesh bit for bit. There are no category verbs, so no chair() or table(): the project's position is that a category is a program written in sand3d, not a function inside it. Units default to millimetres.
The showcase is a front-loading washing machine in 38 lines. Cabinet, drum, door with its glass and handle, detergent drawer and dial are named parts, and four joints say how each one moves.
Exports: GLB, URDF, USDA, IFC, OBJ, STL
The project's GitHub repository documents a single export call where the file extension picks the format and lengths are in metres. GLB carries meshes, materials and the joints in each node's extras at the current pose. URDF gives a link per moving part with joints and limits at rest, with OBJ meshes beside it. USDA rebuilds the same tree in OpenUSD with UsdGeom and UsdPhysics. IFC writes an IFC4 project with an assembly tree, a proxy per part, triangulated face sets and materials. OBJ and STL are the meshes alone.
The caveats are written down. IFC exports the current pose without joints, and its GlobalId values depend on geometry and placement, so editing or moving a part can change an ID and identical parts in separate models can share one. The docs say not to use those IDs to track parts across exports. The library does not write .rvt; for Revit the project points to .ifc, the ISO 16739 standard. .usd, .usdc and .usdz come out of .usda through USD's own usdcat and usdzip.
Setup and pointing a model at it
Python 3.13 is recommended, 3.11 also works, and Blender arrives as the bpy module: 5.2 LTS on 3.13, 4.5 LTS on 3.11, with CI running 5.2. There is no bpy wheel for Python 3.12, nor for Intel Macs from 5.2 on, so those setups use 3.11. Install with pip install sand3d, or uv add sand3d in a uv project.
The LLM side is deliberate rather than incidental. docs/api.md goes into the system prompt with the vocabulary rules, every signature and worked examples, about 3,600 tokens in all, and llms.txt points a model at it. For objects with several parts, the project publishes guidelines with eight rules for assembly programs: plan first, then details, then parts.
A worked example builds a bracket from six parameters: extrude a rounded rectangle into a plate, punch holes on a grid, fillet the top edges, export to GLB. The project also publishes five datasets of sand3d programs on Hugging Face under CC BY 4.0, covering everyday objects, CAD parts, articulated objects, simulation assets and mechanical assemblies.
Limits and open questions
The star and fork counts are a snapshot from one day. The token savings, described as 2 to 6 times fewer than the same part in Blender Python, CadQuery or FeatureScript, are the project's own comparison and the source cites no external benchmark. Nothing in the source names which models were tested. GitHub's listed main language, HTML, sits oddly with a Python library; this likely reflects a docs site or explorer living in the same repository. The citation names Thibault Neveu and Andréa Loy and version 1.6.3.
What it means for building apps with AI
The interesting pattern here is API size as a design constraint. A vocabulary small enough to fit in a prompt, with named handles instead of numbers a model has to compute, is easier to keep on the rails than a large surface with implicit state. Draftship's guide to writing effective prompts covers the same instinct from the app-building side, and sand3d is unaffiliated with Draftship.
Primary sources
- Sandflow-ai/sand3d: 3D modeling in Python, built for LLMs. · GitHub trending AI repositories, Nov 14, 2026
This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.