main
Model: jhlee11/FutureCAD-7B Source: Original Platform
license, language, library_name, pipeline_tag, base_model, datasets, tags
| license | language | library_name | pipeline_tag | base_model | datasets | tags | |||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| apache-2.0 |
|
transformers | text-generation | Qwen/Qwen2.5-7B-Instruct |
|
|
FutureCAD-7B
FutureCAD-7B generates CadQuery programs from text descriptions. It is based on Qwen2.5-7B-Instruct.
Paper · Code · Dataset · BRepGround
Model revisions
| Model | Revision | Training stage |
|---|---|---|
| FutureCAD-7B | main |
Reinforcement learning with GSPO |
| FutureCAD-7B-SFT | sft |
Supervised fine-tuning |
Use revision="sft" to load the SFT model. Omitting revision loads main.
Text-to-CAD
Install the futurecad environment from the code repository.
Run from that repository's root:
python scripts/download_models.py
bash scripts/serve_grounding.sh
In another terminal using the same environment:
export VLLM_USE_FLASHINFER_SAMPLER=0
python -m futurecad.text_to_cad \
--text "Create a cylinder with radius 10 mm and height 20 mm." \
--model weights/futurecad-llm \
--output-dir outputs/cylinder
The command saves the generated Python program and exports STEP/STL when CAD execution succeeds. The repository supplies the prompt template and CadQuery extensions used by the model.
Load with Transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "jhlee11/FutureCAD-7B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype="auto", device_map="auto")
For prompt formatting, generation, execution and evaluation, use the code repository.
Citation
@misc{li2026highfidelitycadgenerationllmdriven,
title={Towards High-Fidelity CAD Generation via LLM-Driven Program Generation and Text-Based B-Rep Primitive Grounding},
author={Jiahao Li and Qingwang Zhang and Qiuyu Chen and Guozhan Qiu and Yunzhong Lou and Xiangdong Zhou},
year={2026},
eprint={2603.11831},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2603.11831},
}
Languages
Jinja
100%