Multi-Modal Foundation · ADSKAILab
Zero-To-CAD Qwen3-VL 2B
Vision-language model generating executable CadQuery CAD programs from multi-view rendered images.
What Zero-To-CAD Qwen3-VL 2B is
Zero-To-CAD Qwen3-VL 2B is an image-to-text model developed by researchers at Autodesk Research (ADSKAILab). It addresses the challenge of reverse engineering 3D geometry from imagery into parametric computer-aided design (CAD) programs. Rather than producing raw meshes or point clouds, it bridges vision-language reasoning and engineering geometry synthesis to produce editable, executable code.
The model takes eight rendered views of a 3D shape and translates them into Python scripts using CadQuery, an open-source parametric CAD modelling library. The resulting programs can be executed directly to generate 3D solids and export standard engineering file formats such as STEP and STL.
For computational researchers and engineers, this provides an interpretable method for parametric 3D reconstruction. The model demonstrates that image-to-CAD synthesis can be acquired using purely synthetic data without relying on real-world engineering design repositories.
How it works
The model is a full fine-tuning of Qwen3-VL-2B-Instruct. It accepts multi-view image inputs—specifically eight 256×256 rendered views (four front, four rear)—alongside conversational textual prompts, and outputs executable CadQuery Python code.
Training was conducted on the synthetic Zero-to-CAD 1M dataset across 979,633 samples without any real-world CAD files. The model was trained using the AdamW optimizer with a cosine learning rate scheduler, a warmup ratio of 0.03, an effective batch size of 16, and bfloat16 mixed precision across 16 NVIDIA H100 GPUs for 3 epochs.
Key facts
- Base model
- Qwen/Qwen3-VL-2B-Instruct
- Fine-tuning type
- Full fine-tuning
- Maximum sequence length
- 4,096 tokens
- Training dataset
- Zero-to-CAD 1M (979,633 synthetic samples)
- Training hardware
- 16 × NVIDIA H100 80GB
- Output format
- Executable CadQuery Python code
What you can use it for
- Parametric Image-to-CAD Reconstruction
- Reconstructing editable, parametric 3D CAD models from eight rendered views of an engineered component.
- Reverse Engineering Pipeline Integration
- Integrating with rendering pipelines to automate end-to-end geometry reconstruction and direct STEP/STL export.
- Program Synthesis Research Baseline
- Serving as an open baseline model for multimodal research focusing on image-to-sequence and CAD program generation.
Run it here
No installation and no GPU of your own: fill in the inputs and it runs on our servers. Results stay in your browser unless you sign in.
Runs on On-demand GPU
Runs on our servers with on-demand compute. A first run needs time to load the model; active capacity can be reused and scales down when idle.Run this model
Zero-To-CAD Qwen3-VL 2B
Strengths and limits
Good at
- Achieves an 82.1% execution success rate and 0.747 mean intersection over union (IoU) on the Zero-to-CAD test benchmark.
- Generates editable and interpretable Python code using CadQuery rather than non-parametric triangle meshes.
- Trained entirely on synthetic data, removing dependencies on proprietary or real-world CAD databases.
Watch out for
- Trained exclusively on synthetic renderings and may struggle when handling photorealistic or noisy real-world images.
- Requires an exact input format of eight clean 256×256 rendered views, with alternative configurations remaining untested.
- Restricted to emitting CadQuery Python code, requiring post-processing or execution steps to derive other CAD formats.
- Complex multi-part assemblies risk exceeding the maximum sequence length limit of 4,096 tokens.
Inputs and outputs
You provide
Image
PNG, JPEG or TIFF. Sent to the model as-is.
Questionoptional
Optional. Without one the model describes the image unprompted.
You get back
The model's completion.
Licence
Check the licence before commercial use · See model card
The See model card licence carries terms worth reading before you depend on it.
Questions
Can I use Zero-To-CAD Qwen3-VL 2B commercially?
Check the licence before commercial use. The See model card licence carries terms worth reading before you depend on it.
How do I run Zero-To-CAD Qwen3-VL 2B without installing anything?
Use the form under “Run it here” on this page: fill in the inputs and press Run model. It runs on our servers on 1× A100 16GB, typically in 12s, and a run costs about 70 credits. Visitors without an account get 1,000 credits a month.
What inputs does the model expect?
The model expects eight rendered views of a 3D shape (four front and four rear) scaled to 256×256 pixels, accompanied by a text prompt requesting CadQuery code.
What is CadQuery?
CadQuery is a Python library used for building parametric 3D CAD models. The model outputs Python scripts that use this library to construct solid geometry.
How is the reconstruction accuracy evaluated?
Performance is measured using voxelized Intersection over Union (IoU) at 64³ resolution against ground-truth solids, calculated across 45-degree rotational increments alongside executable script success rates.
Can the model output standard CAD formats like STEP or STL directly?
The model outputs CadQuery Python code as text; executing that code in a Python environment with CadQuery allows users to export the reconstructed solid as STEP or STL files.
How to cite it
Mohammadmehdi Ataei, Farzaneh Askari, Kamal Rahimi Malekshan, Pradeep Kumar Jayaraman. Zero-to-CAD: Agentic Synthesis of Interpretable CAD Programs at Million-Scale Without Real Data. arXiv:2604.24479, 2026.
Written from Zero-To-CAD Qwen3-VL 2B's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.