---
license: mit
base_model: Qwen/Qwen2.5-Coder-7B-Instruct
tags:
  - text-to-cad
  - cad
  - 3d-modeling
  - parametric
  - lora
  - qwen2
language:
  - en
library_name: transformers
pipeline_tag: text-generation
---

# cad0

**cad0** is a fine-tuned language model for text-to-CAD generation. Given a natural
language description of a mechanical part, it generates [Compact
IR](https://github.com/ecto/vcad) — a token-efficient domain-specific
language for parametric 3D geometry.

## Demo

Try it: [**cad0-demo on HuggingFace
Spaces**](https://huggingface.co/spaces/campedersen/cad0-demo)

For browser/offline use:
[**cad0-mini**](https://huggingface.co/campedersen/cad0-mini) (0.5B, ONNX quantized)

## What It Does

```
Input:  "50x30mm mounting plate with 4 corner holes"
Output: C 50 30 5
        Y 2.5 10
        T 1 5 5 0
        D 0 2
        ...
```

The Compact IR output can be parsed and evaluated by the [vcad
kernel](https://github.com/ecto/vcad) to produce editable parametric
geometry.

## Supported Part Types

| Family | Examples |
|--------|----------|
| Bracket | L-brackets, Z-brackets, mounting plates |
| Standoff | Cylindrical spacers, threaded standoffs |
| Enclosure | Boxes, vented enclosures, rounded shells |
| Gear | Spur gears, hubs |
| Flange | Bolt circles, blind flanges |
| Clip | Snap clips, spring clips |

## Usage

### With Transformers + PEFT

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-Coder-7B-Instruct",
    torch_dtype="auto",
    device_map="auto"
)
model = PeftModel.from_pretrained(base_model, "campedersen/cad0")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-Coder-7B-Instruct")

prompt = "50x30mm mounting plate with 4 corner holes"
messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(messages, tokenize=False,
add_generation_prompt=True)

inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.1,
do_sample=True)
ir = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:],
skip_special_tokens=True)
print(ir)
```

### API Endpoint

```
curl -X POST https://ecto--cad0-training-inference-infer.modal.run \
  -H "Content-Type: application/json" \
  -d '{"prompt": "L-bracket with mounting holes", "temperature": 0.1}'
```

## Compact IR Reference

```
┌──────────────┬──────────────┬──────────────────────────────────┐
│      Op      │    Syntax    │           Description            │
├──────────────┼──────────────┼──────────────────────────────────┤
│ Cube         │ C x y z      │ Box with dimensions x, y, z (mm) │
├──────────────┼──────────────┼──────────────────────────────────┤
│ Cylinder     │ Y r h        │ Cylinder with radius r, height h │
├──────────────┼──────────────┼──────────────────────────────────┤
│ Sphere       │ S r          │ Sphere with radius r             │
├──────────────┼──────────────┼──────────────────────────────────┤
│ Cone         │ K r1 r2 h    │ Cone/frustum                     │
├──────────────┼──────────────┼──────────────────────────────────┤
│ Translate    │ T n x y z    │ Move node n by offset            │
├──────────────┼──────────────┼──────────────────────────────────┤
│ Rotate       │ R n rx ry rz │ Rotate node n (degrees)          │
├──────────────┼──────────────┼──────────────────────────────────┤
│ Scale        │ X n sx sy sz │ Scale node n                     │
├──────────────┼──────────────┼──────────────────────────────────┤
│ Union        │ U a b        │ Boolean union                    │
├──────────────┼──────────────┼──────────────────────────────────┤
│ Difference   │ D a b        │ Boolean subtraction              │
├──────────────┼──────────────┼──────────────────────────────────┤
│ Intersection │ I a b        │ Boolean intersection             │
├──────────────┼──────────────┼──────────────────────────────────┤
│ Shell        │ SH n t       │ Hollow out with wall thickness t │
├──────────────┼──────────────┼──────────────────────────────────┤
│ Fillet       │ F n r        │ Round edges with radius r        │
├──────────────┼──────────────┼──────────────────────────────────┤
│ Chamfer      │ CH n d       │ Bevel edges with distance d      │
└──────────────┴──────────────┴──────────────────────────────────┘
```

Nodes are 0-indexed; each line creates a new node.

## Training

```
┌───────────────────┬───────────────────────────┐
│     Parameter     │           Value           │
├───────────────────┼───────────────────────────┤
│ Base model        │ Qwen2.5-Coder-7B-Instruct │
├───────────────────┼───────────────────────────┤
│ Method            │ QLoRA (4-bit NF4)         │
├───────────────────┼───────────────────────────┤
│ LoRA rank         │ 64                        │
├───────────────────┼───────────────────────────┤
│ LoRA alpha        │ 128                       │
├───────────────────┼───────────────────────────┤
│ Training examples │ 530,531                   │
├───────────────────┼───────────────────────────┤
│ Epochs            │ 1                         │
├───────────────────┼───────────────────────────┤
│ Batch size        │ 64 (effective)            │
├───────────────────┼───────────────────────────┤
│ Learning rate     │ 2e-4 (cosine)             │
├───────────────────┼───────────────────────────┤
│ Hardware          │ 1x NVIDIA H100 80GB       │
├───────────────────┼───────────────────────────┤
│ Training time     │ 9h 15m                    │
└───────────────────┴───────────────────────────┘
```

## Evaluation

```
┌──────────────────────────┬───────────────────────────┐
│          Metric          │           Value           │
├──────────────────────────┼───────────────────────────┤
│ Final eval loss          │ 0.324                     │
├──────────────────────────┼───────────────────────────┤
│ In-distribution accuracy │ 75%                       │
├──────────────────────────┼───────────────────────────┤
│ Out-of-distribution      │ Limited (see Limitations) │
└──────────────────────────┴───────────────────────────┘
```

## Limitations

- Training bias: May add features (holes, fillets) to simple primitives — training
data weighted toward complex manufactured parts
- Dimension ambiguity: Occasionally confuses radius vs diameter for cylinders
- Limited primitives: No native hexagon support; hex shapes require boolean
workarounds
- OOD generalization: Struggles with part types not in training distribution

## Related

- https://github.com/ecto/vcad — Open-source parametric CAD with custom BRep
kernel
- https://vcad.io — Web app with cad0 integration
- https://huggingface.co/campedersen/cad0-mini — 0.5B distilled model for browser
inference

## Citation

@misc{cad0,
  author = {Cam Pedersen},
  title = {cad0: Text-to-CAD Language Model},
  year = {2026},
  publisher = {HuggingFace},
  url = {https://huggingface.co/campedersen/cad0}
}

## License

MIT