List your company
ResearchFormerly WizWand

Accuracy on MMLU (Multi-task Language Understanding)

83.29AccuracyQwen3.5-35B-A3B
22.564438.329754.09569.8603May 26, 2025Jul 30, 2025Oct 3, 2025Dec 8, 2025Feb 11, 2026Apr 17, 2026Jun 22, 2026
Updated 4mo ago

Evaluation results

MethodLinks
Qwen3.5-35B-A3BPrecision=BF162026.0683.29
BaseModel backbone=Qwen3-4B-It, Fine-tuning method=Base2026.0577.6
BaseModel backbone=Qwen3-8B, Fine-tuning method=Base2026.0573.1
BF162026.0273
SVDQuantBit=W8A8, Format=INT2026.0272.9
SVDQuantBit=W8A8, Format=MXFP2026.0272.9
SmoothQuantBit=W8A8, Format=INT2026.0272.8
SmoothQuantBit=W8A8, Format=MXFP2026.0272.8
SVDQuantBit=W8A8, Format=HiF2026.0272.8
RTNBit=W8A8, Format=INT2026.0272.7
RTNBit=W8A8, Format=MXFP2026.0272.7
RTNBit=W8A8, Format=HiF2026.0272.3
SmoothQuantBit=W8A8, Format=HiF2026.0272.3
Qwen3-8BPrecision=BF162026.0672.01
SVDQuantBit=W4A4, Format=NVFP2026.0271.9
SVDQuantBit=W4A4, Format=HiF2026.0271.7
MIXSDModel backbone=Qwen3-8B, Fine-tuning method=MIXSD, Mixing rate λ=0.52026.0571.5
SmoothQuantBit=W4A4, Format=NVFP2026.0270.8
RTNBit=W4A4, Format=NVFP2026.0270.6
RTNBit=W4A4, Format=HiF2026.0270.5
SmoothQuantBit=W4A4, Format=HiF2026.0270.5
GRPOModel=LLaMA-3.2-3B2025.0569
MIXSDModel backbone=Qwen3-4B-It, Fine-tuning method=MIXSD, Mixing rate λ=02026.0568.8
Llama-3.1-8B-InstPrecision=BF162026.0668.77
SVDQuantBit=W4A4, Format=MXFP2026.0268.5
MIXSDModel backbone=Qwen3-4B-It, Fine-tuning method=MIXSD, Mixing rate λ=0.72026.0568.5
SFTModel backbone=Qwen3-4B-It, Fine-tuning method=SFT2026.0568.1
TI-DPOModel=LLaMA-3.2-3B2025.0568
MIXSDModel backbone=Qwen3-8B, Fine-tuning method=MIXSD, Mixing rate λ=0.32026.0568
Phi-4-mini-InstPrecision=BF162026.0667.9
MIXSDModel backbone=Qwen3-8B, Fine-tuning method=MIXSD, Mixing rate λ=02026.0567.8
SFTModel backbone=Qwen3-8B, Fine-tuning method=SFT2026.0567.4
SmoothQuantBit=W4A4, Format=MXFP2026.0267
TPOModel=LLaMA-3.2-3B2025.0567
CPOModel=LLaMA-3.2-3B2025.0566
MIXSDModel backbone=Qwen3-8B, Fine-tuning method=MIXSD, Mixing rate λ=0.72026.0565.5
Llama-3.1-8B (base)Precision=BF162026.0665.39
Qwen2.5-3B-InstPrecision=BF162026.0665.27
RTNBit=W4A4, Format=MXFP2026.0265.2
KTOModel=LLaMA-3.2-3B2025.0565
TDPOModel=LLaMA-3.2-3B2025.0564.5
DPOModel=LLaMA-3.2-3B2025.0564
SIMPOModel=LLaMA-3.2-3B2025.0564
SFTModel=LLaMA-3.2-3B2025.0563
Llama3 8B2026.0262.2
IPOModel=LLaMA-3.2-3B2025.0562
Mistral-7B-InstPrecision=BF162026.0660.87
MIXSDModel backbone=Qwen3-4B-It, Fine-tuning method=MIXSD, Mixing rate λ=0.52026.0559.9
SmoothQuantBit=W8A8, Format=INT2026.0258.8
BaseModel backbone=Qwen3-1.7B, Fine-tuning method=Base2026.0558.5
SmoothQuantBit=W8A8, Format=MXFP2026.0258
SmoothQuantBit=W8A8, Format=HiF2026.0257.8
BF16Format=BF162026.0256.7
RTNBit=W4A4, Format=HiF2026.0256.3
SVDQuantBit=W4A4, Format=HiF2026.0255.9
SVDQuantBit=W8A8, Format=INT2026.0255.7
SVDQuantBit=W8A8, Format=MXFP2026.0255.1
RTNBit=W8A8, Format=INT2026.0254.9
SVDQuantBit=W8A8, Format=HiF2026.0254.3
SmoothQuantBit=W4A4, Format=HiF2026.0254.3
OPSDModel backbone=Qwen3-4B-It, Fine-tuning method=OPSD2026.0553.7
OPSDModel backbone=Qwen3-8B, Fine-tuning method=OPSD2026.0553.6
RTNBit=W8A8, Format=MXFP2026.0253.5
SVDQuantBit=W4A4, Format=NVFP2026.0252.5
COMPOTCR=0.22026.0252.4
SmoothQuantBit=W4A4, Format=MXFP2026.0252.4
RTNBit=W8A8, Format=HiF2026.0252.3
ReplaceMeCR=0.222026.0251.7
SVDQuantBit=W4A4, Format=MXFP2026.0251.3
MIXSDModel backbone=Qwen3-4B-It, Fine-tuning method=MIXSD, Mixing rate λ=0.32026.0550.4
RTNBit=W4A4, Format=MXFP2026.0249
RTNBit=W4A4, Format=NVFP2026.0248.8
SmoothQuantBit=W4A4, Format=NVFP2026.0248.6
Llama2 7BCR=None2025.0940.9
Qwen3-0.6BCompression Ratio=None2026.0240.2
COMPOTCR=0.32026.0239.8
MIXSDModel backbone=Qwen3-1.7B, Fine-tuning method=MIXSD, Mixing rate λ=0.32026.0539.5
SFTModel backbone=Qwen3-1.7B, Fine-tuning method=SFT2026.0534.8
MIXSDModel backbone=Qwen3-1.7B, Fine-tuning method=MIXSD, Mixing rate λ=0.52026.0534.7
SVDQuantBit=W4A4, Format=INT2026.0232.5
ReplaceMeCR=0.312026.0230.6
MIXSDModel backbone=Qwen3-1.7B, Fine-tuning method=MIXSD, Mixing rate λ=02026.0530.4
COMPOTCR=0.42026.0229.6
CoSpaDi (per-layer)CR=0.22025.0928.2
SmoothQuantBit=W4A4, Format=INT2026.0227.5
CoSpaDi (grouped)CR=0.22025.0927.2
COMPOTCompression Ratio=0.32026.0227.2
MIXSDModel backbone=Qwen3-1.7B, Fine-tuning method=MIXSD, Mixing rate λ=0.72026.0527.1
CoSpaDi (per-layer)CR=0.32025.0926.9
SVDQuantBit=W4A4, Format=INT2026.0226.8
CoSpaDiCompression Ratio=0.22026.0226.6
ReplaceMeCR=0.412026.0226.4
COMPOT†Compression Ratio=0.22026.0226.3
COMPOT†Compression Ratio=0.32026.0226.3
MOE+Lngram2026.0526.19
SVD-LLMCompression Ratio=0.22026.0226.1
SVD-LLMCR=0.22025.0925.5
MOE+Engram2026.0525.02
LLM-PrunerCR=0.22026.0225
COMPOTCompression Ratio=0.22026.0224.9