UniParser/MolParser-Mobile

MolParser Mobile

💻 Github | 📄 Report (Coming soon...) | 🚀 Demo

MolParser-Mobile is the next-generation lightweight version of MolParser, designed for ultrafast Optical Chemical Structure Recognition (OCSR). It converts molecular structure images into E-SMILES representations with significantly higher throughput while maintaining state-of-the-art recognition accuracy.

Compared with MolParser 1.0, MolParser-Mobile reduces the model size by 95.4% (216M → 9.98M parameters) and increases inference throughput by 38×, reaching 1,520 molecules/second on a single NVIDIA RTX 4090D GPU. Despite its compact architecture, it achieves better accuracy on multiple real-world OCSR benchmarks, making it suitable for large-scale chemical literature mining and industrial deployment.

🚀 Highlights

• 38× faster than MolParser 1.0 (1,520 Mol/s on RTX 4090D)
• 9.98M parameters (95.4% smaller than MolParser 1.0)
• State-of-the-art accuracy on challenging real-world OCSR benchmarks (BioVista)
• Designed for web-scale chemical literature mining
• Converts molecular images directly into E-SMILES strings

📊 Performance

Model	Parameters	Throughput (RTX 4090D)	Uni-Parser Bench	BioVista	WildMol-10k
Gemma4-31B	31B	OOM	0.073	0.118	0.232
Qwen3.5-397B-A17B	397B	OOM	0.272	0.347	0.303
MolParser 1.0	216M	39.8 Mol/s	0.800	0.703	0.769
MolParser-Mobile	9.98M (-95.4%)	1,520 Mol/s (+3819.1%)	0.823 (+0.023)	0.801 (+0.098)	0.734 (-0.035)

⚡ Usage

Option 1. MolParser Library (Recommended)

The MolParser library provides a convenient interface for molecule detection and recognition.

Clone the repository and install the package:

git clone https://github.com/dptech-corp/MolParser.git
cd MolParser
pip install -e .

Then run:

from molparser import MolParser

parser = MolParser()

result = parser.parse(
    "mol.png",
    rec_only=True,   # Recognition only
)

Option 2. 🤗 Transformers

Load MolParser-Mobile directly with the Hugging Face transformers library.

import torch
from PIL import Image
from transformers import AutoModelForImageTextToText, AutoProcessor

repo_id = "UniParser/MolParser-Mobile"
device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.float16 if device == "cuda" else torch.float32

processor = AutoProcessor.from_pretrained(repo_id, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
    repo_id,
    dtype=dtype,
    trust_remote_code=True,
).to(device).eval()

image = Image.open("mol.png").convert("RGB")
inputs = processor(images=image, return_tensors="pt")
inputs = {k: v.to(device, dtype=dtype) for k, v in inputs.items()}

output_ids = model.generate(**inputs, max_length=256, num_beams=1, do_sample=False)
caption = processor.batch_decode(output_ids, skip_special_tokens=True)[0]
print(caption)

📜 License

MolParser-Mobile Weight

The MolParser-Mobile model weights are provided for non-commercial use only.

For commercial licensing, please contact fangxi@dp.tech or open a discussion on Hugging Face.

MolParser Github Repo

The MolParser library (including E-SMILES post-processing and rendering) is available at https://github.com/dptech-corp/MolParser and is licensed under the Apache License 2.0, which permits commercial use, modification, and distribution, provided that the license and copyright notices are retained.

Note: Model weights, datasets, and third-party dependencies are subject to their respective licenses.

📖 Citation

If you use this datasets in your work, please cite:

@inproceedings{fang2025molparser,
  title={Molparser: End-to-end visual recognition of molecule structures in the wild},
  author={Fang, Xi and Wang, Jiankun and Cai, Xiaochen and Chen, Shangqian and Yang, Shuwen and Tao, Haoyi and Wang, Nan and Yao, Lin and Zhang, Linfeng and Ke, Guolin},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={24528--24538},
  year={2025}
}