A 4B parameter vision-language model fine-tuned from Qwen/Qwen3-VL-4B-Instruct for LaTeX formula recognition from images. This model converts mathematical formula images into accurate LaTeX code.

Model Highlights

Specialized for LaTeX OCR: Trained to accurately transcribe mathematical formulas from images to LaTeX
Multi-Format Support: Handles inline formulas, display equations, matrices, and complex multi-line expressions
High Accuracy: Significantly improved formula recognition over the base model
Vision-Language Architecture: Leverages Qwen3-VL's visual understanding capabilities

Model Description

Property	Value
Base Model	Qwen/Qwen3-VL-4B-Instruct
Model Type	Vision-Language Model (Image-to-Text)
Parameters	4B
Language	English
License	Apache 2.0
Developer	Kassadin88

Training Data

Trained on the LaTeX-OCR dataset for mathematical formula image to LaTeX conversion. The dataset contains rendered LaTeX formulas paired with their source LaTeX code.

Data Composition

Type	Description
Inline formulas	Simple expressions like $E = mc^2$
Display equations	Centered equations with equation numbering
Matrices	Matrix and array environments
Multi-line expressions	Aligned, gathered, and cases environments
Complex formulas	Nested fractions, integrals, summations, and tensor notation

Quick Start

Using Transformers

python

from transformers import AutoModelForCausalLM, AutoProcessor
from PIL import Image

model_name = "Kassadin88/Qwen3-VL-4B-LaTeX-OCR"

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    trust_remote_code=True,
    torch_dtype="auto"
)
processor = AutoProcessor.from_pretrained(
    model_name,
    trust_remote_code=True
)

image = Image.open("formula.png")
messages = [
    {"role": "user", "content": [
        {"type": "image"},
        {"type": "text", "text": "Transcribe the formula in the image to LaTeX."}
    ]}
]

inputs = processor.apply_chat_template(
    messages,
    images=[image],
    return_tensors="pt"
)
outputs = model.generate(**inputs, max_new_tokens=512)
result = processor.decode(outputs[0], skip_special_tokens=True)
print(result)

Using vLLM (Recommended for Production)

bash

vllm serve Kassadin88/Qwen3-VL-4B-LaTeX-OCR \
    --port 8000 \
    --max-model-len 4096 \
    --trust-remote-code

Usage Tips

For Best Results

Use high-resolution, clean images for best recognition accuracy
Crop images tightly around the formula to reduce background noise
For multi-page documents, process one formula at a time

Example Prompt

python

messages = [
    {"role": "user", "content": [
        {"type": "image"},
        {"type": "text", "text": "Convert this mathematical formula to LaTeX."}
    ]}
]

Limitations

May struggle with handwritten formulas or low-quality images
Complex multi-line derivations with mixed text and math may require manual review
Not designed for general OCR tasks (text recognition from documents)
Limited to mathematical notation; does not handle chemical equations or circuit diagrams

Citation

bibtex

@misc{qwen3-vl-4b-latex-ocr,
    author = {Kassadin88},
    title = {Qwen3-VL-4B-LaTeX-OCR: A Fine-Tuned Vision-Language Model for LaTeX OCR},
    year = {2026},
    publisher = {HuggingFace},
    url = {https://huggingface.co/Kassadin88/Qwen3-VL-4B-LaTeX-OCR}
}

Acknowledgments

Base Model: Qwen Team for Qwen3-VL
Training Data: linxy for the LaTeX-OCR dataset

Note: This model is intended for research and educational purposes. Please use responsibly.

clzoro/Qwen3-VL-4B-LaTeX-OCR

Qwen3-VL-4B-LaTeX-OCR