Sandflow Releases 106,000 3D Models With Editable Python Code
Sandflow has released Sand3D, a dataset of 106,000 3D models in which every object is described by Python code built from individual parts. Want a taller chair? Change the leg length, rerun the script, and the model rebuilds itself. The description of the library takes about 3,600 tokens, so it can be pasted into the system prompt of an LLM. This way, language models learn to create and edit 3D objects with commands instead of blindly drawing a mesh.
AI-processed from Habr AI; edited by Hamidun News
Most 3D models used to train neural networks are "frozen" geometry: a set of polygons where you cannot simply say "make the chair back wider." To change an object, you need an artist and an editor. The French startup Sandflow takes a different approach and has released Sand3D, an open dataset of 106,000 3D models in which every object is described by a Python program.
What's inside the dataset
Each model comes with code that builds the object from independent parts. The code lists the object's components, along with their sizes, shapes, positions, and materials. Each part has its own name and function, so the code reads almost like assembly instructions.
Parameters can be changed directly in the script. Take a chair: change the height of the legs or the width of the backrest, run the code again, and you get the rebuilt object with new proportions. Nothing has to be redrawn by hand, and the other parts adjust to the changes.
Besides the code, the authors published images of the models and ready-made OpenUSD files. The models are assembled from several sources: Objaverse, CAD models, Articraft, AutoMate, and Nvidia SimReady. More than 10,000 objects are equipped with moving mechanisms, meaning these are not only static items but also things with hinges and movable joints.
How this relates to language models
The library runs on top of Blender's Python API. It provides common commands for building geometry, fillets, holes, and movable joints. It can be connected to an LLM without fine-tuning: you only need to add the API description to the system prompt. According to the authors, this description takes about 3,600 tokens, which is quite small by the standards of modern context windows.
The idea is simple: if a model can write code, it can "build" a 3D object the same way it writes a function. Instead of generating a point cloud or a mesh, it outputs a script that can then be run, checked, and corrected. The creators themselves suggest using the dataset to train models to create and edit 3D objects. The description-code pair is well suited to this kind of training: an edit becomes a change of a couple of lines rather than regenerating the entire model.
Formats, licenses, and who benefits
The library supports export to several formats:
- GLB, for the web and game engines;
- OpenUSD, for 3D scenes;
- URDF, for robotics simulators;
- OBJ and STL, for general exchange and 3D printing.
URDF support is especially interesting for people working in simulation. An object with movable joints can be loaded straight into a robot simulator and tested, for example, to see how a manipulator opens a cabinet door. However, the authors do not present any ready-made results of such training in the announcement.
The licenses are fairly permissive: the library is released under Apache 2.0, and the data under CC BY 4.0. This means commercial use is allowed, provided the data's authorship is credited. The dataset is available on Hugging Face, and the object collection and code can be viewed on the project website.
What to keep in mind
The limitations should be weighed soberly. The announcement includes no comparisons with other approaches and no quality metrics, so it is unknown how well models trained on this dataset generate new objects. In addition, the dataset was assembled from different sources, and each has its own history of model provenance, so the terms for each part should be checked individually before commercial use.
Still, the benefit for practitioners is clear. For game and web 3D developers, this is a chance to get parametric templates and use them to train an assistant. For those building agents for CAD and design, it offers ready-made examples of "text → code → object." And for roboticists, it provides a large set of objects with moving parts in a format that simulators understand. To test the idea quickly, you only need to paste the 3,600-token API description into a prompt and ask the model to assemble a chair.
- Nov 21, 2026Apex Compute Releases DLSS Alternative for GTX 16 and Radeon: Frame in 1.01 ms vs 1.26 ms
- Nov 21, 2026DeepSeek Weights Show Traces of NVIDIA Blackwell GPUs: 445 of 469 Rows Match on B300
- Nov 21, 2026Amazon SageMaker HyperPod Now Saves Live LLM Requests to S3 at Three Levels
- Nov 21, 2026Amazon Releases SGLang-Neuron: gpt-oss-120b Runs on Trainium Chips Without Changing the API
Key Questions Answered
What's inside the dataset?
Each model comes with code that builds the object from independent parts. The code lists the object's components, along with their sizes, shapes, positions, and materials. Each part has its own name and function, so the code reads almost like assembly instructions. Parameters can be changed directly in the script. Take a chair: change the height of the legs or the width of the backrest, run the code again, and you get the rebuilt object with new proportions.
How this relates to language models?
The library runs on top of Blender's Python API. It provides common commands for building geometry, fillets, holes, and movable joints. It can be connected to an LLM without fine-tuning: you only need to add the API description to the system prompt. According to the authors, this description takes about 3,600 tokens, which is quite small by the standards of modern context windows. The idea is simple: if a model can write code, it can "build" a 3D object the same way it writes a function.
What are the key details on Formats, licenses, and who benefits?
The library supports export to several formats: GLB, for the web and game engines; OpenUSD, for 3D scenes; URDF, for robotics simulators; OBJ and STL, for general exchange and 3D printing. URDF support is especially interesting for people working in simulation. An object with movable joints can be loaded straight into a robot simulator and tested, for example, to see how a manipulator opens a cabinet door. However, the authors do not present any ready-made results of such training in the announcement.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.