Sand3D turns 3D modeling into LLM-friendly code

sand3d is a lightweight Python library built on Blender that lets LLMs generate 3D objects, mechanical parts, and moving joints through concise code.

Instead of making models calculate raw spatial vectors, it treats 3D modeling as a clean programmatic assembly that LLMs can read and write in a few hundred tokens.

What it does

Writing 3D assets in standard Blender scripts, CadQuery, or FeatureScript takes pages of boilerplate code. sand3d operates over Blender's bpy engine but abstracts coordinates away: faces, edges, and parts are referenced by name rather than manually calculated positions.

This makes the code remarkably lean. A given part consumes 2x to 6x fewer tokens than equivalent CAD frameworks. In fact, the entire API specification and examples fit into roughly 3,600 tokens, meaning you can drop the whole reference directly into a system prompt or llms.txt.

It also supports articulation out of the box. A complete front-loading washing machine—including drum, door, handle, detergent drawer, and four moving joints—takes just 38 lines of code.

Exporting is a single function call. You can dump models directly to GLB for game engines, URDF for robot simulators like MuJoCo or Isaac, USDA for Omniverse workflows, and raw OBJ or STL for 3D printing.

Hardware specs and the 100K dataset

The library is open-sourced on GitHub under the Apache 2.0 license, and you can install it via pip install sand3d or uv add sand3d.

It runs on Python 3.13 or 3.11 with Blender bundled as the bpy module (bpy 5.2 LTS on 3.13, 4.5 LTS on 3.11). There are no bpy wheels for Python 3.12, and bpy 5.2 drops support for Intel Macs, so users on those machines should stick to Python 3.11.

Alongside the library, Sandflow AI released a massive dataset of 100,000 3D objects on Hugging Face under CC BY 4.0. Every object has been rewritten as an executable sand3d program.

The collection includes 67,000 everyday assets from Objaverse, 24,000 CAD models, 9,200 articulated props from Articraft, and thousands of industrial parts from NVIDIA SimReady and AutoMate, covering over 10,000 articulated objects and 36,000 joints.

Why it matters

We think sand3d takes the most practical approach yet to LLM-driven 3D generation. Pure diffusion models often produce messy meshes with impossible topologies, while general CAD scripts are too long and brittle for LLM context windows. By turning 3D design into high-level parametric code, sand3d ensures deterministic outputs and easy editing.

Robotics researchers and game developers will get the most value here, especially for generating simulation-ready assets with articulated joints and URDF files. Keep in mind that code-based modeling is built for hard-surface tools, machinery, and furniture; organic shapes like creatures and sculpts still belong in traditional DCC tools.