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

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

Quantization details

• Model type: qwen3_5_moe
• 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.

Model	Quantization	MMLU 1000	Winogrande 300	HumanEval 164	MBPP 300	Average %
Qwen3.6-35B-A3B	Original (71.9GB)	810/1000 81.00%	237/300 79.00%	129/164 78.66%	240/300 80.00%	79.66%
	mlx-lm 4bit (19.5GB)	824/1000 82.40%	229/300 76.33%	151/164 92.07%	243/300 81.00%	82.95%
	UD-MLX-4bit (21.7GB)	821/1000 82.10%	235/300 78.33%	147/164 89.63%	243/300 81.00%	82.77%
	oQ4 (21.1GB)	823/1000 82.30%	232/300 77.33%	152/164 92.68%	237/300 79.00%	82.83%
	oQ4e imatrix (21.1GB)	816/1000 81.60%	236/300 78.67%	154/164 93.90%	244/300 81.33%	83.88%