Opus-MT translation models in ONNX int8 (MediaKit)
31 Opus-MT / Tatoeba-MT models by Helsinki-NLP for translating between English and Russian, German, Spanish, French, Portuguese, Japanese, Polish and Turkish, converted to ONNX and quantized to int8. They are the subtitle translation models of MediaKit, a Windows app: it downloads a folder from here when the user picks a language, checks every file's size and SHA-256, and runs it on the user's computer with ONNX Runtime -- no text leaves the computer.
Nothing in the files is MediaKit-specific; they were tested with ONNX Runtime 1.31.
Files of a model
| File | What it is |
|---|---|
encoder_model.onnx |
the encoder, int8 |
decoder_model_merged.onnx |
one decoder for the first step and the next ones (the KV cache inside an If), int8 |
source.spm, target.spm |
SentencePiece models of the source and the target language, unchanged |
vocab.json |
the shared vocabulary, unchanged |
config.json |
the original Marian configuration (decoder_start_token_id, eos_token_id, ...), unchanged |
LICENSE, NOTICE |
the original model's license; where it comes from and what was changed |
manifest.json lists every file with its size and SHA-256.
Models
4.7 GB in all. Quality: Good -- the best Opus-MT model of the direction by chrF on Tatoeba; Compact -- smaller and faster, 1-4 chrF lower; Experimental -- Japanese, which makes text up on unpunctuated fragments.
| Model | Directions | Converted from | License | Quality | Size |
|---|---|---|---|---|---|
opus-mt-ru-en |
Russian → English | Helsinki-NLP/opus-mt-ru-en | CC BY 4.0 | Good | 147 MB |
opus-mt-en-ru |
English → Russian | Helsinki-NLP/opus-mt-en-ru | Apache-2.0 | Good | 147 MB |
opus-mt-de-en |
German → English | Helsinki-NLP/opus-mt-de-en | Apache-2.0 | Good | 138 MB |
opus-mt-en-de |
English → German | Helsinki-NLP/opus-mt-en-de | CC BY 4.0 | Good | 138 MB |
opus-mt-es-en |
Spanish → English | Helsinki-NLP/opus-mt-es-en | Apache-2.0 | Good | 149 MB |
opus-mt-en-es |
English → Spanish | Helsinki-NLP/opus-mt-en-es | Apache-2.0 | Good | 149 MB |
opus-mt-fr-en |
French → English | Helsinki-NLP/opus-mt-fr-en | Apache-2.0 | Good | 140 MB |
opus-mt-tc-big-en-fr |
English → French | Helsinki-NLP/opus-mt-tc-big-en-fr | CC BY 4.0 | Good | 346 MB |
opus-mt-tc-bible-big-roa-en |
Portuguese → English | Helsinki-NLP/opus-mt-tc-bible-big-roa-en | Apache-2.0 | Good | 354 MB |
opus-mt-tc-big-en-pt |
English → Portuguese | Helsinki-NLP/opus-mt-tc-big-en-pt | CC BY 4.0 | Good | 351 MB |
opus-mt-ja-en |
Japanese → English | Helsinki-NLP/opus-mt-ja-en | Apache-2.0 | Experimental | 142 MB |
opus-tatoeba-en-ja |
English → Japanese | Helsinki-NLP/opus-tatoeba-en-ja | Apache-2.0 | Experimental | 149 MB |
opus-mt-tc-big-zlw-en |
Polish → English | Helsinki-NLP/opus-mt-tc-big-zlw-en | CC BY 4.0 | Good | 362 MB |
opus-mt-tc-bible-big-deu_eng_fra_por_spa-sla |
English → Polish | Helsinki-NLP/opus-mt-tc-bible-big-deu_eng_fra_por_spa-sla | Apache-2.0 | Good | 368 MB |
opus-mt-tr-en |
Turkish → English | Helsinki-NLP/opus-mt-tr-en | Apache-2.0 | Good | 145 MB |
opus-mt-tc-big-en-tr |
English → Turkish | Helsinki-NLP/opus-mt-tc-big-en-tr | CC BY 4.0 | Good | 359 MB |
opus-mt-ROMANCE-en |
Portuguese → English, Spanish → English, French → English | Helsinki-NLP/opus-mt-ROMANCE-en | Apache-2.0 | Compact | 149 MB |
opus-mt-en-ROMANCE |
English → Portuguese, English → Spanish, English → French | Helsinki-NLP/opus-mt-en-ROMANCE | Apache-2.0 | Compact | 149 MB |
opus-mt-pl-en |
Polish → English | Helsinki-NLP/opus-mt-pl-en | Apache-2.0 | Compact | 147 MB |
opus-mt-en-zlw |
English → Polish | Helsinki-NLP/opus-mt-en-zlw | Apache-2.0 | Compact | 138 MB |
opus-mt-en-fr |
English → French | Helsinki-NLP/opus-mt-en-fr | Apache-2.0 | Compact | 140 MB |
opus-mt_tiny_rus-eng |
Russian → English | Helsinki-NLP/opus-mt_tiny_rus-eng | Apache-2.0 | Compact | 38 MB |
opus-mt_tiny_eng-rus |
English → Russian | Helsinki-NLP/opus-mt_tiny_eng-rus | Apache-2.0 | Compact | 38 MB |
opus-mt_tiny_deu-eng |
German → English | Helsinki-NLP/opus-mt_tiny_deu-eng | Apache-2.0 | Compact | 37 MB |
opus-mt_tiny_eng-deu |
English → German | Helsinki-NLP/opus-mt_tiny_eng-deu | Apache-2.0 | Compact | 37 MB |
opus-mt_tiny_spa-eng |
Spanish → English | Helsinki-NLP/opus-mt_tiny_spa-eng | Apache-2.0 | Compact | 37 MB |
opus-mt_tiny_eng-spa |
English → Spanish | Helsinki-NLP/opus-mt_tiny_eng-spa | Apache-2.0 | Compact | 37 MB |
opus-mt_tiny_fra-eng |
French → English | Helsinki-NLP/opus-mt_tiny_fra-eng | Apache-2.0 | Compact | 37 MB |
opus-mt_tiny_eng-fra |
English → French | Helsinki-NLP/opus-mt_tiny_eng-fra | Apache-2.0 | Compact | 37 MB |
opus-mt_tiny_tur-eng |
Turkish → English | Helsinki-NLP/opus-mt_tiny_tur-eng | Apache-2.0 | Compact | 37 MB |
opus-mt_tiny_eng-tur |
English → Turkish | Helsinki-NLP/opus-mt_tiny_eng-tur | Apache-2.0 | Compact | 37 MB |
A multilingual model needs a target-language token at the start of the source text:
| Model | Direction | Token |
|---|---|---|
opus-mt-tc-big-en-pt |
English → Portuguese | >>pob<< (Brazil), >>por<< (Portugal) |
opus-mt-tc-bible-big-deu_eng_fra_por_spa-sla |
English → Polish | >>pol<< |
opus-mt-en-ROMANCE |
English → Portuguese | >>pt_br<< (Brazil), >>pt<< (Portugal) |
opus-mt-en-ROMANCE |
English → Spanish | >>es<< |
opus-mt-en-ROMANCE |
English → French | >>fr<< |
opus-mt-en-zlw |
English → Polish | >>pol<< |
How they were made
Each model: exported to ONNX with Hugging Face Optimum (task text2text-generation-with-past: an encoder and one merged decoder with the KV cache) and quantized to int8 with ONNX Runtime's dynamic quantization (QInt8, MatMulConstBOnly, subgraphs included). The weights were not retrained or fine-tuned. source.spm, target.spm, vocab.json and config.json are copied unchanged.
Toolchain: Python 3.12, torch 2.14.1+cpu, transformers 4.57.6, optimum 2.1.0, optimum-onnx 0.1.0, onnx 1.23.2, onnxruntime 1.31.1, sentencepiece 0.2.2. Another version of the exporter or the quantizer gives other bytes.
Credits and licenses
The models were trained by the Language Technology Research Group at the University of Helsinki (Jörg Tiedemann and the OPUS-MT team) on OPUS data; this repository only converts them. Each folder keeps its original license -- CC BY 4.0 or Apache-2.0, see the table and LICENSE.md.
- Jörg Tiedemann and Santhosh Thottingal. OPUS-MT -- Building open translation services for the World. EAMT 2020.
- Jörg Tiedemann. The Tatoeba Translation Challenge -- Realistic Data Sets for Low Resource and Multilingual MT. WMT 2020.
- Jörg Tiedemann, Mikko Aulamo, Daria Bakshandaeva, Michele Boggia, Stig-Arne Grönroos, Tommi Nieminen, Alessandro Raganato, Yves Scherrer, Raúl Vázquez and Sami Virpioja. Democratizing neural machine translation with OPUS-MT. Language Resources and Evaluation, 2023.