---
pretty_name: CADGUI-Bench
tags:
- gui
- computer-use
- visual-grounding
- solidworks
- benchmark
language:
- zh
size_categories:
- 1K<n<10K
gated: true
extra_gated_heading: Request access
extra_gated_description: Access requests are reviewed manually.
extra_gated_button_content: Request access
extra_gated_prompt: >-
  Please provide your organization or team and intended evaluation.
extra_gated_fields:
  Organization or team: text
  Intended use:
    type: select
    options:
    - Qwen team internal model evaluation
    - Other
---

# CADGUI-Bench

Benchmark data, screenshots, and evaluation code for vision-language models performing SOLIDWORKS GUI tasks.

Access requests are reviewed manually. Include your organization or team and intended evaluation when requesting access.

## Run

```bash
pip install -r requirements.txt
export API_BASE_URL=http://localhost:8000/v1
export API_KEY=EMPTY
python scripts/evaluate.py --task 1.1 --model Qwen3-VL-8B --mode both
```

The `Benchmark/` JSON files reference screenshots under `images/`. `evaluation/benchmark_evaluator.py` contains the scoring utilities, and `scripts/evaluate.py` runs the benchmark against an OpenAI-compatible multimodal endpoint.