OPUS-MT ONNX Model Hub

Mirror of VaishalBusiness/opus (VGT ONNX Model Hub): 1,000+ OPUS-MT translation models in ONNX format for fast inference.

Highlights

• 1,000+ ONNX models from Helsinki-NLP / MarianMT
• Optimized for inference with ONNX Runtime
• Compatible with Hugging Face tokenizers
• Same layout and usage as the original VGT hub

Repository structure

Each model is in its own folder, for example:

Helsinki-NLP-opus-mt-tc-base-bat-zle/
├── config.json
├── decoder_model.onnx
├── decoder_model_merged.onnx
├── decoder_with_past_model.onnx
├── encoder_model.onnx
├── generation_config.json
├── source.spm
├── target.spm
├── special_tokens_map.json
├── tokenizer_config.json
└── vocab.json

• encoder_model.onnx – encoder
• decoder_with_past_model.onnx – decoder with KV cache
• decoder_model_merged.onnx – merged decoder (recommended for speed)
• decoder_model.onnx – base decoder
• source.spm / target.spm, vocab.json – tokenizer files

Usage

Dependencies

pip install huggingface_hub onnxruntime transformers sentencepiece

Load and run a model

Replace Helsinki-NLP-opus-mt-tc-base-bat-zle with any model folder name from the repo.

from huggingface_hub import snapshot_download
import onnxruntime as ort
from transformers import MarianTokenizer
import numpy as np

# Download the model folder (use this repo id after renaming to opus-mt-onnx)
repo_id = "aoiandroid/opus-mt-onnx"  # or aoiandroid/opus if not renamed yet
model_dir = snapshot_download(
    repo_id=repo_id,
    allow_patterns="Helsinki-NLP-opus-mt-tc-base-bat-zle/*",
)

# Load tokenizer
tokenizer = MarianTokenizer.from_pretrained(
    f"{model_dir}/Helsinki-NLP-opus-mt-tc-base-bat-zle"
)

# Encode input
inputs = tokenizer("Hello, how are you?", return_tensors="np")

# Run encoder
enc = ort.InferenceSession(
    f"{model_dir}/Helsinki-NLP-opus-mt-tc-base-bat-zle/encoder_model.onnx"
)
enc_out = enc.run(
    None,
    {
        "input_ids": inputs["input_ids"],
        "attention_mask": inputs["attention_mask"],
    },
)

# Run merged decoder
dec = ort.InferenceSession(
    f"{model_dir}/Helsinki-NLP-opus-mt-tc-base-bat-zle/decoder_model_merged.onnx"
)
decoder_input_ids = np.array([[tokenizer.pad_token_id]], dtype=np.int64)
out = dec.run(
    None,
    {
        "input_ids": decoder_input_ids,
        "encoder_hidden_states": enc_out[0],
    },
)

Attribution

• Original ONNX hub: VaishalBusiness/opus (VGT ONNX Model Hub)
• Underlying models: Helsinki-NLP OPUS-MT / MarianMT
• ONNX conversions and hosting follow the original project’s intent; licenses of each model are unchanged.

License

Each model keeps its original license. When using a model, cite the original authors and comply with their license and attribution requirements. This repository only provides a mirror of ONNX conversions.