kaojai/quickmt-en-th-onnx

QuickMT en-th ONNX Q8

This package runs the QuickMT en-th translation model in a browser with a custom ONNX Runtime Web decoder loop. It is not compatible with generic pipeline("translation") loading because the exported decoder has no KV cache.

Provenance

• Source model: quickmt/quickmt-en-th
• Source revision: a4f1466054c237018111d97074b07e8006e6e2ca
• EOLE source revision: 735a720561be7d9c7dd74fd4397a630a28d459a8
• Quantization: dynamic per-channel Q8
• License: CC BY 4.0

Files

• tokenizer.json and tokenizer_config.json: source tokenizer
• target-tokenizer/: target tokenizer
• onnx/encoder_model_quantized.onnx: Q8 encoder
• onnx/decoder_model_merged_quantized.onnx: Q8 decoder
• release-manifest.json: byte counts and SHA-256 values for this release

Limitations

The first download is approximately 230 MiB per language direction. The decoder does not use a KV cache, so long outputs require repeated prefix decoding. Use revision-pinned URLs and CacheStorage in the browser; do not bundle the model in an application repository.