English注册

vanch007/Hy-MT2-30B-A3B-oQ6e-fp16

来源：🤗translationapache-2.030.1B 参数激活 3B60 GBsafetensors✓ 5 个校验和18 天前更新
或提交种子

Hy-MT2-30B-A3B-oQ6e-fp16

MLX safetensors conversion of tencent/Hy-MT2-30B-A3B, quantized with oMLX's enhanced oQ6 quantizer (6-bit affine with group size 64, FP16 scales and activations, 8-bit shared experts).

Model Details

  • Base Architecture: Hunyuan V3 MoE (48 hidden layers, 128 routed experts with top-8 routing, 1 shared expert, 2048 hidden size, 120,832 vocab size).
  • Format: MLX safetensors (5 shards, 24.62 GB).
  • Quantization: oQ6e (6-bit affine gs64, 8-bit shared experts, FP16 non-quantized weights).
  • Data-driven Calibration: Source-specific oQe imatrix calibrated over multilingual code & text (128 samples x 512 tokens).

oMLX Usage

Place this repository into an oMLX model directory and refresh the model list.

License

Apache-2.0, inherited from the base model.