kaptionai/hy-mt2-ane

Hy-MT2 1.8B for the Apple Neural Engine (Kaption)

A Core ML conversion of tencent/Hy-MT2-1.8B (revision 9a341cd1b679d3efd23b46e847b01745a71ed792), packaged for on-device translation on the Apple Neural Engine in the Kaption desktop app for macOS 15 and later on Apple silicon.

This is an independent adaptation: Tencent did not make it and does not endorse Kaption.

License and attribution

Licensed under the Apache License 2.0, the license of the source model.

• Source model: Tencent Hy-MT2-1.8B, https://huggingface.co/tencent/Hy-MT2-1.8B, revision 9a341cd1b679d3efd23b46e847b01745a71ed792, model.safetensors SHA-256 29e9117a44c79f81857613601968ff482d8a23c2d6736a1710bba9e5ca4762e5. Copyright (C) 2026 Tencent. LICENSE.txt is the source model's license file, unchanged.
• Changes made by Kaption: converted to Core ML, split into chunks and compressed (see below). The weights are no longer bit-identical to the source.

Files

File	Weights	SHA-256	Bytes
kaption-ane-mt-pal4g16-c4p.zip	4-bit, int8 embedding	2516e99baed28cfd58dfbdfbbab0c397c48488b53b3ce3674c2fc270a2075be0	1,058,815,797
kaption-ane-mt-pal6g16-c4.zip	6-bit, fp16 embedding	781dfafabeaa0d5314a35193aca2e386c3f1341c398284f81ba6e515f2f3ecfe	1,713,714,548

Each zip holds four transformer chunks (chunk0..3.mlpackage, each with a KV cache in Core ML state), the split LM head (head.mlpackage), the token embedding (embed.*), the tokenizer (tokenizer.bin, byte-level BPE from the source tokenizer.json), mt.json and kaption-ane-package.json, which lists every file with its SHA-256 and size. Manifest SHA-256: 4bec8cdb4fab5cb3c6f66ae7ed803dcf79d270e7a7c72c98c397e80eafd724f7 (4-bit), 0c32885d17748a256d1d4d0587fba8d474a624a9625f8c2ef803a9eb165bdc81 (6-bit).

How it was converted

1. Weights loaded from the source model.safetensors above into an fp32 PyTorch model with the attention rewritten for the Neural Engine layout (numerically equivalent), a fixed 512-token context and up to 4 decoding lanes.
2. Converted with coremltools 9.0 (torch 2.7.0) to ML Programs (minimum deployment target macOS 15), then palettized per grouped channel (group size 16) to 4 or 6 bits with coremltools.optimize.coreml.

On an Apple M1 (16 GB), against the full-precision model on 70 synthetic requests in 5 directions: chrF 85.6 (4-bit) and 93.5 (6-bit); a short message takes about 0.47 s one at a time. Kaption checks every download against the SHA-256 and size above and qualifies the compiled model on the device before using it; anything that fails falls back to Kaption's CPU model or Apple Translation.

Use

This package is made for Kaption's kaption-ane helper and its protocol; it is not a drop-in Transformers checkpoint. To use the model elsewhere, start from tencent/Hy-MT2-1.8B.