Part of the Hy-MT2 MLX collection.

Hy-MT2-1.8B-8bit (MLX)

Apple MLX weights for Tencent-Hunyuan/Hy-MT2-1.8B, Tencent Hunyuan's multilingual translation model.

TL;DR

Variant	1.8B 8-bit
Best for	smaller 1.8B checkpoint
Runtime	mlx-lm
Official code	Tencent-Hunyuan/Hy-MT2
MLX code	ailuntx/Hy-MT2-MLX
Hardware	Apple Silicon recommended; HF Spaces CPU fallback is only a load smoke test

Quick Start

pip install mlx-lm
PROMPT=$'Translate the following text into English. Note that you should only output the translated result without any additional explanation:\n\n今天天气真好。'
mlx_lm.generate --model mlx-community/Hy-MT2-1.8B-8bit \
  --prompt "$PROMPT" \
  --max-tokens 128 \
  --temp 0.7 \
  --top-p 0.6 \
  --top-k 20

For conversion and staging tools:

git clone https://github.com/ailuntx/Hy-MT2-MLX.git
cd Hy-MT2-MLX
python scripts/infer_mlx.py --model /path/to/mlx/checkpoint --text "今天天气真好。" --target-lang English

Variants

Variant	Best for
Hy-MT2-1.8B-bfloat16	high-quality 1.8B baseline
Hy-MT2-1.8B-8bit	smaller 1.8B checkpoint
Hy-MT2-1.8B-4bit	smallest 1.8B checkpoint
Hy-MT2-7B-bfloat16	high-precision 7B conversion
Hy-MT2-7B-8bit	7B size/quality middle ground
Hy-MT2-7B-4bit	smallest 7B checkpoint

Conversion Notes

Component	Source	MLX handling
model weights	official dense Hy-MT2 checkpoint	converted with MLX/mlx-lm tooling
tokenizer/config	official checkpoint	copied through for mlx-lm loading
quantized variants	bfloat16 MLX baseline	derived with MLX quantization settings

Validation

Local Apple Silicon is the intended runtime. The Hy-MT2 HF Space starts and loads the model on Linux CPU fallback, but cpu-basic can exceed request timeouts even for very small generation tests.

License

License follows the upstream Tencent Hunyuan release.

Citation

@misc{hymt2-mlx,
  title  = {Hy-MT2-MLX: Apple MLX port of Hy-MT2},
  author = {ailuntx},
  year   = {2026},
  url    = {https://github.com/ailuntx/Hy-MT2-MLX},
}

@misc{zheng2026hymt2familyfastefficient,
  title         = {Hy-MT2: A Family of Fast, Efficient and Powerful Multilingual Translation Models in the Wild},
  author        = {Mao Zheng and Zheng Li and Tao Chen and Bo Lv and Mingrui Sun and Mingyang Song and Jinlong Song and Hong Huang and Decheng Wu and Hai Wang and Yifan Song and Yanfeng Chen and Guanwei Zhang},
  year          = {2026},
  eprint        = {2605.22064},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2605.22064},
}