Language Modeling on WikiText-2
1.08Perplexity (PPL)Calibration-free SLEB
Evaluation results
| Method | Links | |||||||||||||||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Calibration-free SLEBBase Model=Qwen3-8B, Budget=800, Evals=666, # Removed=362026.05 | 1.08 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| TCN-SEQ-IBackbone=GP-VAE, tok/s=~ 90002025.12 | 1.61 | — | — | — | — | — | — | — | — | — | — | — | — | — | 0.4773 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| TCN-SEQ-CBackbone=GP-VAE, tok/s=~ 90002025.12 | 1.66 | — | — | — | — | — | — | — | — | — | — | — | — | — | 0.5077 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| TCN-SEQ-BBackbone=GP-VAE, tok/s=~ 89002025.12 | 1.73 | — | — | — | — | — | — | — | — | — | — | — | — | — | 0.5455 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| TCN-SEQ-JBackbone=GP-VAE, tok/s=~ 90002025.12 | 1.76 | — | — | — | — | — | — | — | — | — | — | — | — | — | 0.5635 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| TCN-SEQ-KBackbone=GP-VAE, tok/s=~ 90002025.12 | 1.82 | — | — | — | — | — | — | — | — | — | — | — | — | — | 0.6005 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| RTNAverage Bitwidth (bits)=2.1252026.03 | 2 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Full PrecisionBackbone=Llama3-70B, Mixed-Precision=false, Bits=16, Quantizer Type=Uniform2026.02 | 2.59 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| UnprunedBackbone=Llama-3.1-70B, Sparsity=0%2026.04 | 2.79 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| FP16Model Family=LLaMA-3, Model Size=70B, Bits=162024.06 | 2.86 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| FP16W Quant.=None, Backbone=LLaMA-3-70B2025.11 | 2.86 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Pretrained DenseWeight Update=None, Sparsity=0, Model Architecture=LLaMA-3, Model Size=70B2026.02 | 2.86 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| BaselineBackbone=LLaMA3-70B, Precision=FP162026.05 | 2.87 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| BaselineDifferential Privacy=None2025.12 | 2.95 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 62.55 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SA-ADP2025.12 | 3.05 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 60.96 | 0.7697 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Pretrained DenseWeight Update=None, Sparsity=0, Model Architecture=LLaMA-2, Model Size=70B2026.02 | 3.12 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DenseModel=LLaMA-2 70B, Sparsity (%)=0.002025.09 | 3.319 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 0 | — | — | — | — | — | — | — | — | — | — | |
| FP16Model Family=LLaMA-2, Model Size=70B, Bits=162024.06 | 3.32 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| FP16W Quant.=None, Backbone=LLaMA-2-70B2025.11 | 3.32 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| FP16Model Family=Llama-2, Model Size=70B, Precision=FP162023.06 | 3.32 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DenseSparsity pattern=Unstructured (50%), Model family=LLaMA-2, Model size=70B2026.06 | 3.32 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DenseSparsity pattern=2:4 Pruning, Model family=LLaMA-2, Model size=70B2026.06 | 3.32 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| FP16Avg. Bit=16, Backbone=LLaMA2-70B2024.10 | 3.32 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CMPQAvg. Bit=4, Backbone=LLaMA2-70B2024.10 | 3.39 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| AWQModel Family=Llama-2, Model Size=70B, Quantization=INT4, Group Size=1282023.06 | 3.41 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| AWQAvg. Bit=4, Backbone=LLaMA2-70B2024.10 | 3.41 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| GPTQModel Family=Llama-2, Model Size=70B, Quantization=INT4, Group Size=1282023.06 | 3.42 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| GPTQ-RModel Family=Llama-2, Model Size=70B, Quantization=INT4, Group Size=1282023.06 | 3.43 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| RTNModel Family=Llama-2, Model Size=70B, Quantization=INT4, Group Size=1282023.06 | 3.46 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| TORQBackbone=LLaMA3-70B2026.05 | 3.49 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| FP16Model Family=LLaMA, Model Size=65B, Precision=FP162023.06 | 3.53 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| QuIPAvg. Bit=4, Backbone=LLaMA2-70B2024.10 | 3.53 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DenseModel=LLaMA 65B, Sparsity (%)=0.002025.09 | 3.531 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 0 | — | — | — | — | — | — | — | — | — | — | |
| SGDModel Architecture=RoBERTa2026.04 | 3.547 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DenseSparsity pattern=Unstructured (50%), Model family=LLaMA, Model size=65B2026.06 | 3.56 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DenseSparsity pattern=2:4 Pruning, Model family=LLaMA, Model size=65B2026.06 | 3.56 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| GPTQAvg. Bit=4, Backbone=LLaMA2-70B2024.10 | 3.59 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| AWQModel Family=LLaMA, Model Size=65B, Quantization=INT4, Group Size=1282023.06 | 3.62 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ShiftAddLLMBits=3, Model Family=LLaMA-2, Model Size=70B, Group Size=1282024.06 | 3.64 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SingleQuantW Quant.=RTN, Backbone=LLaMA-2-70B2025.11 | 3.65 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| GPTQModel Family=LLaMA, Model Size=65B, Quantization=INT4, Group Size=1282023.06 | 3.66 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| GPTQ-RModel Family=LLaMA, Model Size=65B, Quantization=INT4, Group Size=1282023.06 | 3.66 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| EMP-MagnitudeModel=LLaMA-2 70B, Sparsity (%)=36.662025.09 | 3.662 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 0.343 | — | — | — | — | — | — | — | — | — | — | |
| RTNModel Family=LLaMA, Model Size=65B, Quantization=INT4, Group Size=1282023.06 | 3.67 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| RTNAvg. Bit=4, Backbone=LLaMA2-70B2024.10 | 3.67 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ScaleBITS + RTNBackbone=Llama3-70B, Mixed-Precision=true, Bits=3.1, Quantizer Type=Uniform2026.02 | 3.69 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SpinQuantW Quant.=GPTQ, Backbone=LLaMA-2-70B2025.11 | 3.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CMPQAvg. Bit=3, Backbone=LLaMA2-70B2024.10 | 3.7 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| ShiftAddLLM (Lat.)Model Family=LLaMA-2, Model Size=70B, Bits=3, Group Size=1282024.06 | 3.72 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| AWQModel Family=LLaMA-2, Model Size=70B, Bits=3, Group Size=1282024.06 | 3.74 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| AWQModel Family=Llama-2, Model Size=70B, Quantization=INT3, Group Size=1282023.06 | 3.74 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| AWQAvg. Bit=3, Backbone=LLaMA2-70B2024.10 | 3.74 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| AdaCubicModel Architecture=RoBERTa2026.04 | 3.756 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| OmniQuantBits=3, Model Family=LLaMA-2, Model Size=70B, Group Size=1282024.06 | 3.78 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| MergeQuantW Quant.=GPTQ, Backbone=LLaMA-2-70B2025.11 | 3.78 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| QuaRotW Quant.=GPTQ, Backbone=LLaMA-2-70B2025.11 | 3.79 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DuQuantW Quant.=RTN, Backbone=LLaMA-2-70B2025.11 | 3.79 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Ours (CVR+EC)Weight Update=Yes, Sparsity=50%, Model Architecture=LLaMA-2, Model Size=70B2026.02 | 3.81 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SpinQuantW Quant.=RTN, Backbone=LLaMA-2-70B2025.11 | 3.82 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| EMP-WandaModel=LLaMA-2 70B, Sparsity (%)=40.322025.09 | 3.821 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 0.502 | — | — | — | — | — | — | — | — | — | — | |
| FP16/BF16Backbone=Mixtral-8x7B, Quantization=FP16/BF162026.05 | 3.84 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| QuIPAvg. Bit=3, Backbone=LLaMA2-70B2024.10 | 3.85 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Wanda+ECWeight Update=Yes, Sparsity=50%, Model Architecture=LLaMA-2, Model Size=70B2026.02 | 3.86 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| GPTQ-RModel Family=Llama-2, Model Size=70B, Quantization=INT3, Group Size=1282023.06 | 3.86 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| EMP-MagnitudeModel=LLaMA 65B, Sparsity (%)=36.612025.09 | 3.865 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 0.334 | — | — | — | — | — | — | — | — | — | — | |
| OPTQModel Family=LLaMA-2, Model Size=70B, Bits=3, Group Size=1282024.06 | 3.88 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Pretrained DenseWeight Update=None, Sparsity=0, Model Architecture=Qwen2.5, Model Size=72B2026.02 | 3.88 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| GPTQModel Family=Llama-2, Model Size=70B, Quantization=INT3, Group Size=1282023.06 | 3.88 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| GuidedQuantBackbone=Llama3-70B, Mixed-Precision=false, Bits=3, Quantizer Type=Non-uniform2026.02 | 3.9 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| AWQModel Family=LLaMA, Model Size=65B, Quantization=INT3, Group Size=1282023.06 | 3.95 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| WandaWeight Update=No, Sparsity=50%, Model Architecture=LLaMA-2, Model Size=70B2026.02 | 3.98 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| BaselineModel=Phi-3.5-MoE, W-bit / GS=FP16, Time (s)=-2026.04 | 3.98 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| RTNModel Family=Llama-2, Model Size=70B, Quantization=INT3, Group Size=1282023.06 | 3.98 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| LUT-GEMMModel Family=LLaMA-2, Model Size=70B, Bits=3, Group Size=1282024.06 | 4.01 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| WandaModel=LLaMA-2 70B, Sparsity (%)=50.002025.09 | 4.026 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 0.707 | — | — | — | — | — | — | — | — | — | — | |
| AWQBackbone=Mixtral-8x7B, Quantization=W4A162026.05 | 4.03 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| MiLoModel=Mixtral-8x7B, Avg.bit=3.42026.05 | 4.03 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| EMP-WandaModel=LLaMA 65B, Sparsity (%)=39.992025.09 | 4.06 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 0.529 | — | — | — | — | — | — | — | — | — | — | |
| SlimLLM-g128Backbone=Llama3-70B, Mixed-Precision=true, Bits=3.1, Quantizer Type=Uniform2026.02 | 4.08 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| CRePESparsity pattern=Unstructured (50%), Model family=LLaMA-2, Model size=70B2026.06 | 4.09 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| FP16Model Family=LLaMA, Model Size=30B, Precision=FP162023.06 | 4.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| FP16/BF16Backbone=LLaMA-30B, Quantization=FP16/BF162026.05 | 4.1 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| DenseModel=LLaMA 30B, Sparsity (%)=0.002025.09 | 4.101 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 0 | — | — | — | — | — | — | — | — | — | — | |
| RIASparsity pattern=Unstructured (50%), Model family=LLaMA-2, Model size=70B2026.06 | 4.11 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| TILEQ_sModel=Mixtral-8x7B, Avg.bit=3.162026.05 | 4.12 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| QuaRotW Quant.=RTN, Backbone=LLaMA-2-70B2025.11 | 4.14 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| BaselineModel=Mixtral-8x7B, W-bit / GS=FP16, Time (s)=–2026.04 | 4.14 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| GPTQModel Family=LLaMA, Model Size=65B, Quantization=INT3, Group Size=1282023.06 | 4.17 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| GPTQ-RModel Family=LLaMA, Model Size=65B, Quantization=INT3, Group Size=1282023.06 | 4.21 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| AWQModel Family=LLaMA, Model Size=30B, Quantization=INT4, Group Size=1282023.06 | 4.21 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| AWQBackbone=LLaMA-30B, Quantization=W4A162026.05 | 4.21 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| GPTQ-RModel Family=LLaMA, Model Size=30B, Quantization=INT4, Group Size=1282023.06 | 4.22 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| WandaSparsity pattern=Unstructured (50%), Model family=LLaMA-2, Model size=70B2026.06 | 4.22 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| RTNModel Family=LLaMA, Model Size=30B, Quantization=INT4, Group Size=1282023.06 | 4.23 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Ours (CVR+EC)Weight Update=Yes, Sparsity=4:8, Model Architecture=LLaMA-2, Model Size=70B2026.02 | 4.24 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| RTNModel Family=LLaMA, Model Size=65B, Quantization=INT3, Group Size=1282023.06 | 4.24 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| GPTQModel Family=LLaMA, Model Size=30B, Quantization=INT4, Group Size=1282023.06 | 4.24 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SparseGPTSparsity pattern=Unstructured (50%), Model family=LLaMA-2, Model size=70B2026.06 | 4.25 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| WandaModel=LLaMA 65B, Sparsity (%)=50.002025.09 | 4.267 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | 0.736 | — | — | — | — | — | — | — | — | — | — | |
| FP16Model Backbone=LLaMA2-13B2025.12 | 4.356 | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — | — |