Modelpublic

gatesushi/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-oQ4-mtp

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes24downloads

Model Card

Based on nightmedia's Qwen3.6-27B-Architect-Polaris2-Fable-B-F451

This quantization was performed using omlx (0.5.4 dev). It utilizes mtp and oQ quantization.

The Model appears to perform better in benchmarks and has seemed effective in chats.

Here are the benchmarks executed in omlx. These are 100 tests with the think set to off.

Model: Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-oQ4-mtp

Benchmark         Accuracy    Time(s)
--------------------------------------------------------------
MMLU                 90.0%     145.3
MMLU_PRO             73.0%      65.8
HELLASWAG            94.0%      64.6
TRUTHFULQA           91.0%      44.6
ARC_CHALLENGE        91.0%      41.2
WINOGRANDE           85.0%      39.1
GSM8K                96.0%     308.7
MATHQA               59.0%      44.7


Model: Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-mxfp4-mlx
Benchmark         Accuracy    Time(s)
--------------------------------------------------------------
MMLU                 88.0%     116.4
MMLU_PRO             77.0%        59
HELLASWAG            93.0%      62.2
TRUTHFULQA           90.0%      43.1
ARC_CHALLENGE        92.0%      39.5
WINOGRANDE           82.0%      37.4
GSM8K                96.0%     338.4
MATHQA               52.0%      44.8


Model: Qwen3.6-27B-Miraculix-mxfp8-mlx
Benchmark         Accuracy    Time(s)
--------------------------------------------------------------
MMLU                 82.0%     127.1
MMLU_PRO             73.0%      63.4
HELLASWAG            93.0%      65.5
TRUTHFULQA           91.0%      48.3
ARC_CHALLENGE        92.0%      45.7
WINOGRANDE           83.0%      43.9
GSM8K                96.0%     344.7
MATHQA               54.0%        54

Model: Qwen3.6-27B-Fable-Fusion-711-BigBubba-717-mxfp8-mlx
Benchmark         Accuracy    Time(s)
--------------------------------------------------------------
MMLU                 89.0%     158.2
MMLU_PRO             75.0%      74.2
HELLASWAG            93.0%      72.9
TRUTHFULQA           93.0%      52.7
ARC_CHALLENGE        92.0%      49.1
WINOGRANDE           83.0%        47
GSM8K                97.0%     454.8
MATHQA               55.0%      60.9


Model: Qwen3.6-27B-ElCHUD-mxfp8-mlx
Benchmark         Accuracy    Time(s)
--------------------------------------------------------------
MMLU                 88.0%       171
MMLU_PRO             75.0%        78
HELLASWAG            93.0%      74.2
TRUTHFULQA           92.0%        53
ARC_CHALLENGE        92.0%      49.5
WINOGRANDE           84.0%        47
GSM8K                96.0%     447.8
MATHQA               55.0%      55.1

Model: Qwen3.6-27B-Akka-mxfp8-mlx
Benchmark         Accuracy    Time(s)
--------------------------------------------------------------
MMLU                 84.0%     132.5
MMLU_PRO             72.0%      75.1
HELLASWAG            92.0%      74.4
TRUTHFULQA           92.0%      53.4
ARC_CHALLENGE        91.0%      50.9
WINOGRANDE           83.0%      49.9
GSM8K                97.0%     425.7
MATHQA               53.0%        57