Language Generation on WikiText2
2.9PerplexityBaseline
Evaluation Results
| Method | Links | |
|---|---|---|
| BaselinePrec.=FP16, Model=3-70B2026.05 | 2.9 | |
| BaselinePrec.=FP16, Model=2-70B2026.05 | 3.31 | |
| W16A16 BaselineBits=W16A16, Model=Llama2-70B2026.03 | 3.33 | |
| QuIPPrec.=W4A16, Model=3-70B2026.05 | 3.4 | |
| SliderQuantBits=W4A16, Model=Llama2-70B2026.03 | 3.41 | |
| OSAQ+GPTQPrec.=W4A16, Model=3-70B2026.05 | 3.42 | |
| CBQBits=W4A16, Model=Llama2-70B2026.03 | 3.46 | |
| OmniQuantBits=W4A16, Model=Llama2-70B2026.03 | 3.47 | |
| OmniQuantPrec.=W4A16, Model=2-70B2026.05 | 3.47 | |
| OSAQ+GPTQPrec.=W4A16, Model=2-70B2026.05 | 3.48 | |
| W16A16Precision=W16A16, Backbone=LLaMA-1-65B2024.02 | 3.53 | |
| QuIPPrec.=W4A16, Model=2-70B2026.05 | 3.53 | |
| OSAQ+AWQPrec.=W4A16, Model=2-70B2026.05 | 3.53 | |
| AWQPrec.=W4A16, Model=2-70B2026.05 | 3.57 | |
| OSAQ+AWQPrec.=W4A16, Model=3-70B2026.05 | 3.57 | |
| GPTQBits=W4A16, Model=Llama2-70B2026.03 | 3.58 | |
| MagRPrec.=W4A16, Model=2-70B2026.05 | 3.58 | |
| GPTQPrec.=W4A16, Model=2-70B2026.05 | 3.58 | |
| AWQBits=W4A16, Model=Llama2-70B2026.03 | 3.6 | |
| GPTQPrec.=W4A16, Model=3-70B2026.05 | 3.6 | |
| RTNBits=W4A16, Model=Llama2-70B2026.03 | 3.67 | |
| RTNPrec.=W4A16, Model=2-70B2026.05 | 3.67 | |
| AWQPrec.=W4A16, Model=3-70B2026.05 | 3.7 | |
| OSAQ+AWQPrec.=W3A16 g128, Model=2-70B2026.05 | 3.75 | |
| OmniQuantPrec.=W3A16 g128, Model=2-70B2026.05 | 3.78 | |
| OSAQ+GPTQPrec.=W3A16 g128, Model=2-70B2026.05 | 3.79 | |
| OSAQ+QuIPPrec.=W3A16, Model=2-70B2026.05 | 3.81 | |
| AWQPrec.=W3A16 g128, Model=2-70B2026.05 | 3.83 | |
| GPTQPrec.=W3A16 g128, Model=2-70B2026.05 | 3.85 | |
| QuIPPrec.=W3A16, Model=2-70B2026.05 | 3.85 | |
| OmniQuantPrec.=W3A16, Model=2-70B2026.05 | 3.92 | |
| MagRPrec.=W3A16 g128, Model=2-70B2026.05 | 3.95 | |
| RTNPrec.=W3A16 g128, Model=2-70B2026.05 | 3.97 | |
| W16A16Precision=W16A16, Backbone=LLaMA-1-30B2024.02 | 4.1 | |
| OSAQ+GPTQPrec.=W3A16, Model=2-70B2026.05 | 4.21 | |
| OSAQ+AWQPrec.=W3A16 g128, Model=3-70B2026.05 | 4.61 | |
| MagRPrec.=W3A16, Model=2-70B2026.05 | 4.64 | |
| OSAQ+QuIPPrec.=W3A16, Model=3-70B2026.05 | 4.65 | |
| QuIPPrec.=W3A16, Model=3-70B2026.05 | 4.7 | |
| AWQPrec.=W3A16 g128, Model=3-70B2026.05 | 4.8 | |
| GPTQPrec.=W3A16, Model=2-70B2026.05 | 4.82 | |
| DB-LLMPrecision=W2A16, Group Size=64, Backbone=LLaMA-1-65B2024.02 | 4.84 | |
| W16A16 BaselineBits=W16A16, Model=Llama2-13B2026.03 | 4.88 | |
| BaselinePrec.=FP16, Model=2-13B2026.05 | 4.88 | |
| OSAQ+GPTQPrec.=W3A16 g128, Model=3-70B2026.05 | 4.99 | |
| SliderQuantBits=W4A16, Model=Llama2-13B2026.03 | 5 | |
| QuIPPrec.=W4A16, Model=2-13B2026.05 | 5.01 | |
| OmniQuantBits=W4A16, Model=Llama2-13B2026.03 | 5.02 | |
| CBQBits=W4A16, Model=Llama2-13B2026.03 | 5.02 | |
| OmniQuantPrec.=W4A16, Model=2-13B2026.05 | 5.02 | |
| OSAQ+AWQPrec.=W4A16, Model=2-13B2026.05 | 5.04 | |
| OSAQ+GPTQPrec.=W4A16, Model=2-13B2026.05 | 5.04 | |
| W16A16Precision=W16A16, Backbone=LLaMA-1-13B2024.02 | 5.09 | |
| AWQBits=W4A16, Model=Llama2-13B2026.03 | 5.12 | |
| AWQPrec.=W4A16, Model=2-13B2026.05 | 5.12 | |
| GPTQBits=W4A16, Model=Llama2-13B2026.03 | 5.13 | |
| GPTQPrec.=W4A16, Model=2-13B2026.05 | 5.13 | |
| MagRPrec.=W4A16, Model=2-13B2026.05 | 5.17 | |
| RTNBits=W4A16, Model=Llama2-13B2026.03 | 5.2 | |
| RTNPrec.=W4A16, Model=2-13B2026.05 | 5.2 | |
| GPTQPrec.=W3A16 g128, Model=3-70B2026.05 | 5.2 | |
| AWQPrecision=W3A16, Backbone=LLaMA-1-65B2024.02 | 5.21 | |
| OSAQ+AWQPrec.=W3A16 g128, Model=2-13B2026.05 | 5.23 | |
| OSAQ+QuIPPrec.=W3A16, Model=2-13B2026.05 | 5.25 | |
| OmniQuantPrec.=W3A16 g128, Model=2-13B2026.05 | 5.28 | |
| W16A16 BaselineBits=W16A16, Model=Qwen2.5-14B2026.03 | 5.3 | |
| OSAQ+GPTQPrec.=W3A16 g128, Model=2-13B2026.05 | 5.3 | |
| AWQPrec.=W3A16 g128, Model=2-13B2026.05 | 5.32 | |
| QuIPPrec.=W3A16, Model=2-13B2026.05 | 5.34 | |
| GPTQPrec.=W3A16 g128, Model=2-13B2026.05 | 5.42 | |
| MagRPrec.=W3A16 g128, Model=2-13B2026.05 | 5.45 | |
| W16A16 BaselineBits=W16A16, Model=Llama2-7B2026.03 | 5.47 | |
| BaselinePrec.=FP16, Model=2-7B2026.05 | 5.47 | |
| Full-precision#Bits=16/16, Size (GB)=13.542026.05 | 5.47 | |
| OSAQ+GPTQPrec.=W3A16, Model=3-70B2026.05 | 5.49 | |
| RTNPrec.=W3A16 g128, Model=2-13B2026.05 | 5.51 | |
| DB-LLMPrecision=W2A16, Group Size=64, Backbone=LLaMA-1-30B2024.02 | 5.52 | |
| MagR†Prec.=W2A16 g128, Model=2-70B2026.05 | 5.52 | |
| OmniQuantPrec.=W3A16, Model=2-13B2026.05 | 5.58 | |
| SliderQuantBits=W4A16, Model=Llama2-7B2026.03 | 5.61 | |
| OmniQuantPrecision=W2A16, Group Size=64, Backbone=LLaMA-1-65B2024.02 | 5.65 | |
| CBQBits=W4A16, Model=Llama2-7B2026.03 | 5.67 | |
| W16A16Precision=W16A16, Backbone=LLaMA-1-7B2024.02 | 5.68 | |
| OSAQ+GPTQPrec.=W3A16, Model=2-13B2026.05 | 5.72 | |
| OSAQ+GPTQPrec.=W4A16, Model=2-7B2026.05 | 5.73 | |
| OmniQuantBits=W4A16, Model=Llama2-7B2026.03 | 5.74 | |
| OmniQuantPrec.=W4A16, Model=2-7B2026.05 | 5.74 | |
| SliderQuantBits=W4A16, Model=Qwen2.5-14B2026.03 | 5.8 | |
| GPTQBits=W4A16, Model=Llama2-7B2026.03 | 5.83 | |
| CBQBits=W4A16, Model=Qwen2.5-14B2026.03 | 5.83 | |
| GPTQPrec.=W4A16, Model=2-7B2026.05 | 5.83 | |
| GPTQPrec.=W3A16, Model=3-70B2026.05 | 5.88 | |
| MagRPrec.=W4A16, Model=2-7B2026.05 | 5.91 | |
| AWQPrec.=W3A16, Model=3-70B2026.05 | 5.92 | |
| OmniQuantBits=W4A16, Model=Qwen2.5-14B2026.03 | 5.94 | |
| QuIPPrec.=W4A16, Model=2-7B2026.05 | 5.94 | |
| OSAQ+GPTQ†Prec.=W2A16 g128, Model=2-70B2026.05 | 5.97 | |
| OSAQ+AWQPrec.=W4A16, Model=2-7B2026.05 | 5.99 | |
| OmniQuantPrec.=W3A16 g128, Model=2-7B2026.05 | 6.03 | |
| DuQuant + NBC#Bits=4/4, Size (GB)=4.972026.05 | 6.03 |