Image: GitHub
Why it mattersAn LLM that writes 3D code has to fit the library's whole API into its context. A library that fits in 3,600 tokens is one a model can use without a retrieval step.
Sandflow-ai released Sand3d on 3 November 2026, an open-source Python library for 3D modeling built for language models to use. In three days the repository reached 331 stars, and it sits at rank 12 on Trendshift's daily list. The description in the repo header: "3D modeling in Python, built for LLMs." Apache-2.0 licence.
What Sand3d is
Sand3d is a thin layer on top of Blender's bpy module. It gives a flat API, s3d.verb, across six families of verbs. The six are profiles, solids and paths, edits, frames, articulation and assembly. Profiles include rect, circle, slot, polygon and arc. Solids include extrude, revolve, sweep, box and sphere. Edits include cut, join, hole and fillet. Articulation gives joint, pose and seat. About 60 verbs in total.
The design point, in Sandflow's own words: the whole API fits in about 3,600 tokens. Sandflow says writing a part in Sand3d takes "2 to 6x fewer tokens" than in Blender Python or CadQuery. The library runs on Python 3.11 or 3.13 with bpy 4.5 or 5.2.
What it exports
Exports cover five formats. GLB for the web. URDF for robotics simulators. USDA for Nvidia Omniverse and other OpenUSD tools. OBJ and STL for general 3D software and 3D printing. Sandflow also published five datasets on Hugging Face under CC BY 4.0. They cover household objects, mechanical parts, articulated assemblies and simulation-ready assets.
Why the token count matters
The 3,600-token figure is the one worth sitting with. A language model asked to write code against a library has to carry the library's API in its context. If the API runs to tens of thousands of tokens of reference, the model needs a retrieval step: look up the right verbs, then write the code. If the API fits in 3,600 tokens, it travels in the system prompt beside the user's task. The model can write a part on the first pass with no retrieval.
That is not a benchmark against another tool. It is a design choice Sandflow made about how to host a Python API an agent can use. CadQuery and Blender Python are both larger than this: a model using them has to sample from more surface and is more likely to hallucinate a method name. A smaller surface area, with examples in the README, is a lower bar for an agent to clear.
Who it is for
Two groups will test this in the next week. A design team using an LLM to prototype 3D parts from text: the Sand3d README shows a 38-line washing machine with named parts and four joints. A roboticist generating URDF from a description: the joint and pose verbs do work a one-shot image generator cannot do. The joints have to make physical sense. The quality of the output in both cases is a function of the model, not the library. What the library gives both is a surface small enough to fit in a single context window.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.