SevenAIGmbH/opus-mt-de-pl

https://huggingface.co/Helsinki-NLP/opus-mt-de-pl with ONNX weights to be compatible with Transformers.js.

Same layout as the Xenova/opus-mt-* repos: fp32 graphs plus fp16, quantized (q8), int8, uint8, q4, q4f16 and bnb4 variants under onnx/. Converted with the transformers.js 3.7.0 conversion scripts (optimum + onnxruntime 1.20.1), with these fixes:

• decoder_model_merged is re-merged without onnxslim (the slimmed graph fails in onnxruntime's extended graph optimizations for the past-key-values branch) and saved as ONNX IR 8, so it loads in onnxruntime-web 1.14 (@xenova/transformers v2).
• The tied lm_head weight Transpose is placed after the embedding Gather, so 8-bit quantization does not keep an extra fp32 copy of the embedding matrix.

Usage (Transformers.js)

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

const translator = await pipeline('translation', 'SevenAIGmbH/opus-mt-de-pl');
const output = await translator('Hallo, wie geht es dir?');

Attribution

The model weights are by the Language Technology Research Group at the University of Helsinki (OPUS-MT), released under the Apache License 2.0 as declared in the source repository. Only the format was changed (converted to ONNX and quantized).