CoolFace
Modelpublic

Jundot/Qwen3.6-35B-A3B-oQ4e-mtp

sourceHugging Faceupdated 3mo agoView on Hugging Face
23likes2.9kdownloads
Model Card

Qwen3.6-35B-A3B-oQ4e-mtp

This model was quantized using oQ (oMLX v0.4.5.dev1) mixed-precision quantization.

Quantization details

  • —Model type: qwen35moe
  • —Bits: 4
  • —Group size: 64
  • —Format: MLX safetensors

Benchmark

All runs used thinking off. Evaluation used the oMLX standard benchmark runner with greedy decoding, so the same model artifacts reproduce the same results. Disk size is shown under each quantization label. Original is shown as a reference and is excluded from bold best-score marking. Bold values mark the best quantized score within each model for each benchmark, including the average; ties are all bolded.

<table> <thead> <tr> <th width="16%">Model</th> <th width="20%">Quantization</th> <th>MMLU 1000</th> <th>Winogrande 300</th> <th>HumanEval 164</th> <th>MBPP 300</th> <th>Average %</th> </tr> </thead> <tbody> <tr> <td rowspan="5"><strong>Qwen3.6-35B-<wbr>A3B</strong></td> <td>Original<br>(71.9GB)</td> <td align="right">810/1000<br>81.00%</td> <td align="right">237/300<br>79.00%</td> <td align="right">129/164<br>78.66%</td> <td align="right">240/300<br>80.00%</td> <td align="right">79.66%</td> </tr> <tr> <td>mlx-lm 4bit<br>(19.5GB)</td> <td align="right"><strong>824/1000<br>82.40%</strong></td> <td align="right">229/300<br>76.33%</td> <td align="right">151/164<br>92.07%</td> <td align="right">243/300<br>81.00%</td> <td align="right">82.95%</td> </tr> <tr> <td>UD-MLX-4bit<br>(21.7GB)</td> <td align="right">821/1000<br>82.10%</td> <td align="right">235/300<br>78.33%</td> <td align="right">147/164<br>89.63%</td> <td align="right">243/300<br>81.00%</td> <td align="right">82.77%</td> </tr> <tr> <td>oQ4<br>(21.1GB)</td> <td align="right">823/1000<br>82.30%</td> <td align="right">232/300<br>77.33%</td> <td align="right">152/164<br>92.68%</td> <td align="right">237/300<br>79.00%</td> <td align="right">82.83%</td> </tr> <tr> <td><strong>oQ4e imatrix<br>(21.1GB)</strong></td> <td align="right">816/1000<br>81.60%</td> <td align="right"><strong>236/300<br>78.67%</strong></td> <td align="right"><strong>154/164<br>93.90%</strong></td> <td align="right"><strong>244/300<br>81.33%</strong></td> <td align="right"><strong>83.88%</strong></td> </tr> </tbody> </table>