OptimaMT/en-hi

OptimaMT en → hi (English to Hindi)

Drop-in copy of Xenova/opus-mt-en-hi at revision 7aad72006a15 with a lighter merged decoder: ONNX Runtime's fast mode (disable_quant_qdq = 1, 2.3–2.5× faster) at half the memory. Same weights, same tokens. Part of OptimaMT: all 49 pairs, named OptimaMT/<src>-<tgt>.

Used in production:

• OpenCircle, a desktop screen translator for Windows and Linux, ships these models since v0.1.4.

Memory after load	Linux	Windows	Identical tokens
disable_quant_qdq = 1	522 → 279 MiB (-47 %)	498 → 255 MiB (-49 %)	yes
ONNX Runtime defaults	279 → 218 MiB (-22 %)	196 → 164 MiB (-16 %)	yes

FLORES-200 devtest (1,012 sentences): the same translations as the original, BLEU 11.33, chrF 32.0 (greedy).

ONNX Runtime 1.31.1, CPU, 4 threads (Linux; Windows in a VM). Interactive results: OptimaMT/results · method: REPORT.md.

Use

import { pipeline } from "@huggingface/transformers";

const translator = await pipeline("translation", "OptimaMT/en-hi", {
  dtype: "q8",
  session_options: { extra: { session: { disable_quant_qdq: "1" } } },
});
const [out] = await translator("The lighthouse stands on the coast.");
console.log(out.translation_text);

Python (translate.py: onnxruntime, sentencepiece, numpy):

python translate.py OptimaMT/en-hi "The lighthouse stands on the coast."

Files

Only onnx/decoder_model_merged_quantized.onnx is rebuilt; every other file is the original, byte for byte. Vocabulary: 61,950 tokens.

Licence and credits

• Weights: OPUS-MT by the University of Helsinki, CC BY 4.0.
• ONNX export: Xenova/opus-mt-en-hi at revision 7aad72006a15.
• Modified: the decoder's output projection was moved out of its two branches (graph change only). CC BY 4.0.