List your company

Physical Reasoning on PIQA (Accuracy)

82.5AccuracyAAAC
49.781658.275866.7775.2642Jul 1, 2025Aug 23, 2025Oct 16, 2025Dec 9, 2025Jan 31, 2026Mar 26, 2026May 19, 2026

Evaluation results

MethodLinks
AAACModel=Qwen3.5-27B, Group Size (g)=162026.0582.5
RTNModel=Qwen3.5-27B, Group Size (g)=162026.0582.3
IF4Model=Qwen3.5-27B, Group Size (g)=162026.0582.2
BF16Model=Qwen3.5-27B2026.0582.1
AWQModel=Qwen3.5-27B, Group Size (g)=162026.0581.8
RSPruning ratio=25%, Backbone=Qwen2-57B-A14B, Evaluation protocol=Zero-shot, Calibration samples=100, Calibration dataset=Zyda22025.0781.34
Qwen2-57B-A14BPruning ratio=0%, Backbone=Qwen2-57B-A14B, Evaluation protocol=Zero-shot, Calibration samples=100, Calibration dataset=Zyda22025.0781.23
MoonlightPruning ratio=0%, Backbone=Moonlight, Evaluation protocol=Zero-shot, Calibration samples=100, Calibration dataset=Zyda22025.0781.12
FLAPPruning ratio=25%, Backbone=Qwen2-57B-A14B, Evaluation protocol=Zero-shot, Calibration samples=100, Calibration dataset=Zyda22025.0781.12
MoNEPruning ratio=25%, Backbone=Qwen2-57B-A14B, Evaluation protocol=Zero-shot, Calibration samples=100, Calibration dataset=Zyda22025.0781.07
RSPruning ratio=25%, Backbone=Moonlight, Evaluation protocol=Zero-shot, Calibration samples=100, Calibration dataset=Zyda22025.0781.01
NHA-Qwen3-30BA3BModel=NHA-Qwen3-30BA3B, Architecture Type=Native Hybrid Attention, Full-attention layers=32025.1080.96
MoNEPruning ratio=25%, Backbone=Moonlight, Evaluation protocol=Zero-shot, Calibration samples=100, Calibration dataset=Zyda22025.0780.85
MC-SMoEPruning ratio=25%, Backbone=Moonlight, Evaluation protocol=Zero-shot, Calibration samples=100, Calibration dataset=Zyda22025.0780.36
Qwen3-30B-A3BPruning ratio=0%, Backbone=Qwen3-30B-A3B, Evaluation protocol=Zero-shot, Calibration samples=100, Calibration dataset=Zyda22025.0780.3
BaselineModel Architecture=Qwen3-30B-A3B2026.0580.3
Deepseek-V2-LitePruning ratio=0%, Backbone=Deepseek-V2-Lite, Evaluation protocol=Zero-shot, Calibration samples=100, Calibration dataset=Zyda22025.0780.25
MC-SMoEPruning ratio=25%, Backbone=Qwen2-57B-A14B, Evaluation protocol=Zero-shot, Calibration samples=100, Calibration dataset=Zyda22025.0780.2
RSPruning ratio=25%, Backbone=Deepseek-V2-Lite, Evaluation protocol=Zero-shot, Calibration samples=100, Calibration dataset=Zyda22025.0780.14
IF4Model=Qwen3.5-9B, Group Size (g)=162026.0580
AWQModel=Qwen3.5-9B, Group Size (g)=162026.0580
AAACModel=Qwen3.5-9B, Group Size (g)=162026.0580
RSPruning ratio=25%, Backbone=Qwen3-30B-A3B, Evaluation protocol=Zero-shot, Calibration samples=100, Calibration dataset=Zyda22025.0779.92
BF16Model=Qwen3.5-9B2026.0579.9
OLMoEPruning ratio=0%, Backbone=OLMoE, Evaluation protocol=Zero-shot, Calibration samples=100, Calibration dataset=Zyda22025.0779.76
MoNEPruning ratio=25%, Backbone=Deepseek-V2-Lite, Evaluation protocol=Zero-shot, Calibration samples=100, Calibration dataset=Zyda22025.0779.76
MC-SMoEPruning ratio=25%, Backbone=Qwen3-30B-A3B, Evaluation protocol=Zero-shot, Calibration samples=100, Calibration dataset=Zyda22025.0779.71
Qwen3-30BA3BModel=Qwen3-30BA3B, Architecture Type=Transformer2025.1079.49
MoNEPruning ratio=25%, Backbone=Qwen3-30B-A3B, Evaluation protocol=Zero-shot, Calibration samples=100, Calibration dataset=Zyda22025.0779.27
RTNModel=Qwen3.5-9B, Group Size (g)=162026.0579.2
FLAPPruning ratio=25%, Backbone=Deepseek-V2-Lite, Evaluation protocol=Zero-shot, Calibration samples=100, Calibration dataset=Zyda22025.0778.18
MoNEPruning ratio=25%, Backbone=OLMoE, Evaluation protocol=Zero-shot, Calibration samples=100, Calibration dataset=Zyda22025.0778.07
FLAPPruning ratio=25%, Backbone=Moonlight, Evaluation protocol=Zero-shot, Calibration samples=100, Calibration dataset=Zyda22025.0777.97
SSQR-2%Avg. Bitwidth=4.765, Backbone=Qwen3-8B2025.0777.97
TORQModel Architecture=Qwen3-30B-A3B2026.0577.93
BF16 BaselineAvg. Bitwidth=16.000, Backbone=Qwen3-8B2025.0777.8
GPTQAvg. Bitwidth=4.125, Backbone=Qwen3-8B2025.0777.58
HPTQAvg. Bitwidth=3.125, Backbone=Qwen3-8B2025.0777.53
SSQR-3%Avg. Bitwidth=5.085, Backbone=Qwen3-8B2025.0777.53
HPTQAvg. Bitwidth=4.125, Backbone=Qwen3-8B2025.0777.42
SSQR-5%Avg. Bitwidth=5.725, Backbone=Qwen3-8B2025.0777.42
FLAPPruning ratio=25%, Backbone=Qwen3-30B-A3B, Evaluation protocol=Zero-shot, Calibration samples=100, Calibration dataset=Zyda22025.0777.31
SSQR-3%Avg. Bitwidth=4.085, Backbone=Qwen3-8B2025.0777.31
SSQR-4%Avg. Bitwidth=5.405, Backbone=Qwen3-8B2025.0777.31
SSQR-1%Avg. Bitwidth=4.445, Backbone=Qwen3-8B2025.0777.04
HRTNAvg. Bitwidth=4.125, Backbone=Qwen3-8B2025.0776.99
SSQR-5%Avg. Bitwidth=4.725, Backbone=Qwen3-8B2025.0776.71
SSQR-4%Avg. Bitwidth=4.405, Backbone=Qwen3-8B2025.0776.61
SSQR-2%Avg. Bitwidth=3.765, Backbone=Qwen3-8B2025.0776.22
HRTNAvg. Bitwidth=3.125, Backbone=Qwen3-8B2025.0776.12
SOLARTraining Stage=SFT-1B2025.1275.9
MIDUS-HMLTraining Stage=SFT-1B2025.1275.9
RTNAvg. Bitwidth=4.125, Backbone=Qwen3-8B2025.0775.9
SSQR-1%Avg. Bitwidth=3.445, Backbone=Qwen3-8B2025.0775.84
MC-SMoEPruning ratio=25%, Backbone=Deepseek-V2-Lite, Evaluation protocol=Zero-shot, Calibration samples=100, Calibration dataset=Zyda22025.0775.57
OpT-DeUSTraining Stage=SFT-1B2025.1275.52
BaseTraining Stage=SFT-1B2025.1275.46
Llama ProTraining Stage=SFT-1B2025.1275.46
GPTQAvg. Bitwidth=3.125, Backbone=Qwen3-8B2025.0775.46
AngularPruning ratio=25%, Backbone=Qwen3-30B-A3B, Evaluation protocol=Zero-shot, Calibration samples=100, Calibration dataset=Zyda22025.0775.3
SOLARTraining Stage=CPT-1B2025.1275.24
LESATraining Stage=SFT-1B2025.1275.24
OSTQuantModel Architecture=Qwen3-30B-A3B2026.0575.09
BaseTraining Stage=CPT-1B2025.1275.08
MIDUS-HMLTraining Stage=CPT-1B2025.1275.08
Avg-DeUSTraining Stage=SFT-1B2025.1274.97
FLAPPruning ratio=25%, Backbone=OLMoE, Evaluation protocol=Zero-shot, Calibration samples=100, Calibration dataset=Zyda22025.0774.81
LESATraining Stage=CPT-1B2025.1274.59
Llama ProTraining Stage=CPT-1B2025.1274.54
OpT-DeUSTraining Stage=CPT-1B2025.1274.54
Avg-DeUSTraining Stage=CPT-1B2025.1274.27
SSQR-5%Avg. Bitwidth=3.725, Backbone=Qwen3-8B2025.0773.23
SSQR-4%Avg. Bitwidth=3.405, Backbone=Qwen3-8B2025.0772.8
HPTQAvg. Bitwidth=2.125, Backbone=Qwen3-8B2025.0772.52
AngularPruning ratio=25%, Backbone=OLMoE, Evaluation protocol=Zero-shot, Calibration samples=100, Calibration dataset=Zyda22025.0771.65
AngularPruning ratio=25%, Backbone=Deepseek-V2-Lite, Evaluation protocol=Zero-shot, Calibration samples=100, Calibration dataset=Zyda22025.0771.33
QuaRotModel Architecture=Qwen3-30B-A3B2026.0571.11
RTNAvg. Bitwidth=3.125, Backbone=Qwen3-8B2025.0770.89
AngularPruning ratio=25%, Backbone=Qwen2-57B-A14B, Evaluation protocol=Zero-shot, Calibration samples=100, Calibration dataset=Zyda22025.0770.02
SSQR-3%Avg. Bitwidth=3.085, Backbone=Qwen3-8B2025.0769.59
AngularPruning ratio=25%, Backbone=Moonlight, Evaluation protocol=Zero-shot, Calibration samples=100, Calibration dataset=Zyda22025.0768.5
MC-SMoEPruning ratio=25%, Backbone=OLMoE, Evaluation protocol=Zero-shot, Calibration samples=100, Calibration dataset=Zyda22025.0767.19
SSQR-2%Avg. Bitwidth=2.765, Backbone=Qwen3-8B2025.0766.76
RSPruning ratio=25%, Backbone=OLMoE, Evaluation protocol=Zero-shot, Calibration samples=100, Calibration dataset=Zyda22025.0766.16
HRTNAvg. Bitwidth=2.125, Backbone=Qwen3-8B2025.0765.78
GPTQModel Architecture=Qwen3-30B-A3B2026.0562.81
SSQR-1%Avg. Bitwidth=2.445, Backbone=Qwen3-8B2025.0761.64
GPTQAvg. Bitwidth=2.125, Backbone=Qwen3-8B2025.0757.83
RTNAvg. Bitwidth=2.125, Backbone=Qwen3-8B2025.0751.63
RTNModel Architecture=Qwen3-30B-A3B2026.0551.04