List your company
ResearchFormerly WizWand

Language Modeling on Wiki

2Perplexity (PPL)Wanda
1.4145.36959.32513.2805Oct 2, 2025Nov 4, 2025Dec 8, 2025Jan 10, 2026Feb 13, 2026Mar 18, 2026Apr 21, 2026
Updated 6mo ago

Evaluation results

MethodLinks
WandaWall-clock (h)=0.159, #GPUs=1, Model=LLaMA-2-7B, Sparsity=90%2025.102
DenseSparsity=0%, Backbone=LLaMA-2-7B2025.105.47
SAFEModel=LLaMA-2-13B, Sparsity=50%2025.105.73
ALPSModel=LLaMA-2-13B, Sparsity=50%2025.105.9
L-ADMMModel=LLaMA-2-13B, Sparsity=50%2025.105.92
WandaModel=LLaMA-2-13B, Sparsity=50%2025.105.97
SparseGPTModel=LLaMA-2-13B, Sparsity=50%2025.106.03
ELSAModel=LLaMA-2-13B, Sparsity=50%2025.106.54
SAFEModel=LLaMA-2-7B, Sparsity=50%2025.106.72
L-ADMMModel=LLaMA-2-7B, Sparsity=50%2025.106.8
MagnitudeModel=LLaMA-2-13B, Sparsity=50%2025.106.83
ALPSModel=LLaMA-2-7B, Sparsity=50%2025.106.86
SAFEModel=LLaMA-2-13B, Sparsity=60%2025.106.9
WandaModel=LLaMA-2-7B, Sparsity=50%2025.106.92
SparseGPTModel=LLaMA-2-7B, Sparsity=50%2025.107.01
DenseModel=Llama-3.1-8B, Calibration Method=Baseline2026.037.03
AWQ (W4A16)Model=Llama-3.1-8B, Calibration Method=COLA2026.037.21
SparseLLMModel=LLaMA-2-7B, Sparsity=50%2025.107.23
AWQ (W4A16)Model=Llama-3.1-8B, Calibration Method=ZipCal2026.037.31
AWQ (W4A16)Model=Llama-3.1-8B, Calibration Method=ZipCal (Multi-Domain)2026.037.32
ELSAModel=LLaMA-2-7B, Sparsity=50%2025.107.5
Wanda (25%)Model=Llama-3.1-8B, Calibration Method=ZipCal (Multi-Domain)2026.037.52
ALPSModel=LLaMA-2-13B, Sparsity=60%2025.107.56
L-ADMMModel=LLaMA-2-13B, Sparsity=60%2025.107.57
Wanda (25%)Model=Llama-3.1-8B, Calibration Method=ZipCal2026.037.6
Wanda (25%)Model=Llama-3.1-8B, Calibration Method=COLA2026.037.62
DenseModel=Gemma-2-9B, Calibration Method=Baseline2026.037.63
GPTQ (W4A16)Model=Llama-3.1-8B, Calibration Method=COLA2026.037.71
GPTQ (W4A16)Model=Llama-3.1-8B, Calibration Method=ZipCal2026.037.71
Wanda (25%)Model=Gemma-2-9B, Calibration Method=ZipCal (Multi-Domain)2026.037.86
Wanda (25%)Model=Gemma-2-9B, Calibration Method=COLA2026.037.87
Wanda (25%)Model=Gemma-2-9B, Calibration Method=ZipCal2026.037.92
GPTQ (W4A16)Model=Gemma-2-9B, Calibration Method=COLA2026.037.93
GPTQ (W4A16)Model=Gemma-2-9B, Calibration Method=ZipCal2026.037.93
ELSAModel=LLaMA-2-13B, Sparsity=60%2025.107.96
SAFESparsity=4:8, Backbone=LLaMA-2-7B2025.108.043
ALPSSparsity=4:8, Backbone=LLaMA-2-7B2025.108.103
L-ADMMSparsity=4:8, Backbone=LLaMA-2-7B2025.108.12
SparseGPTModel=LLaMA-2-13B, Sparsity=60%2025.108.27
WandaModel=LLaMA-2-13B, Sparsity=60%2025.108.4
GPTQ (W4A16)Model=Gemma-2-9B, Calibration Method=ZipCal (Multi-Domain)2026.038.44
SparseGPTSparsity=4:8, Backbone=LLaMA-2-7B2025.108.508
FP16Model=Qwen-2.5-3B-Instruct2026.048.56
GPTQ (W4A16)Model=Llama-3.1-8B, Calibration Method=ZipCal (Multi-Domain)2026.038.59
WandaSparsity=4:8, Backbone=LLaMA-2-7B2025.108.603
SparseLLMSparsity=4:8, Backbone=LLaMA-2-7B2025.108.679
SAFEModel=LLaMA-2-7B, Sparsity=60%2025.109.02
ELSAModel=LLaMA-2-7B, Sparsity=60%2025.109.16
ELSASparsity=4:8, Backbone=LLaMA-2-7B2025.109.2
ALPSModel=LLaMA-2-7B, Sparsity=60%2025.109.33
L-ADMMModel=LLaMA-2-7B, Sparsity=60%2025.109.4
SAFESparsity=2:4, Backbone=LLaMA-2-7B2025.109.914
ALPSSparsity=2:4, Backbone=LLaMA-2-7B2025.109.945
ELSASparsity=2:4, Backbone=LLaMA-2-7B2025.1010.15
L-ADMMSparsity=2:4, Backbone=LLaMA-2-7B2025.1010.19
SparseGPTModel=LLaMA-2-7B, Sparsity=60%2025.1010.2
SAFEModel=LLaMA-3.2-3B, Sparsity=50%2025.1010.68
SparseLLMModel=LLaMA-2-7B, Sparsity=60%2025.1010.74
WandaModel=LLaMA-2-7B, Sparsity=60%2025.1010.79
SparseGPTSparsity=2:4, Backbone=LLaMA-2-7B2025.1010.87
ALPSModel=Gemma-2-2B, Sparsity=50%2025.1010.93
L-ADMMModel=Gemma-2-2B, Sparsity=50%2025.1011.02
ELSAModel=LLaMA-2-13B, Sparsity=70%2025.1011.14
SparseLLMSparsity=2:4, Backbone=LLaMA-2-7B2025.1011.29
ALPSModel=LLaMA-3.2-3B, Sparsity=50%2025.1011.31
L-ADMMModel=LLaMA-3.2-3B, Sparsity=50%2025.1011.56
SparseGPTModel=Gemma-2-2B, Sparsity=50%2025.1011.58
SAFEModel=Gemma-2-2B, Sparsity=50%2025.1011.61
MeSHModel Backbone=Pythia-6.9B, Scheme=Recursive (-31.25%), Layers=6+10R2+6, Variant=+mesh2025.1011.64
2SSP (25%)Model=Gemma-2-9B, Calibration Method=ZipCal2026.0311.68
MagnitudeModel=LLaMA-2-13B, Sparsity=60%2025.1011.82
ELSAModel=LLaMA-3.2-3B, Sparsity=50%2025.1012.06
WandaModel=Gemma-2-2B, Sparsity=50%2025.1012.07
WandaSparsity=2:4, Backbone=LLaMA-2-7B2025.1012.13
Recursive baselineModel Backbone=Pythia-6.9B, Scheme=Recursive (-31.25%), Layers=6+10R2+6, Variant=base2025.1012.14
SparseGPTModel=LLaMA-3.2-3B, Sparsity=50%2025.1012.27
SAFEModel=LLaMA-2-13B, Sparsity=70%2025.1012.47
2SSP (25%)Model=Gemma-2-9B, Calibration Method=COLA2026.0312.56
WandaModel=LLaMA-3.2-3B, Sparsity=50%2025.1013.01
ELSAModel=Gemma-2-2B, Sparsity=50%2025.1013.05
ELSAModel=LLaMA-2-7B, Sparsity=70%2025.1013.2
MeSHModel Backbone=Pythia-2.8B, Scheme=Recursive (-31.25%), Layers=6+10R2+6, Variant=+mesh2025.1013.6
2SSP (25%)Model=Gemma-2-9B, Calibration Method=ZipCal (Multi-Domain)2026.0314.14
ALPSModel=LLaMA-2-13B, Sparsity=70%2025.1014.17
Recursive baselineModel Backbone=Pythia-2.8B, Scheme=Recursive (-31.25%), Layers=6+10R2+6, Variant=base2025.1014.18
ALPSModel=Gemma-2-2B, Sparsity=60%2025.1014.42
L-ADMMModel=Gemma-2-2B, Sparsity=60%2025.1014.65
L-ADMMModel=LLaMA-2-13B, Sparsity=70%2025.1014.81
SAFEModel=Gemma-2-2B, Sparsity=60%2025.1015.22
MeSHModel Size=Pythia-1.4B, Scheme=Recursive, Layers=4+8R2+4, Variant=+mesh, Parameter reduction percentage=-33.3%2025.1015.84
MagnitudeSparsity=4:8, Backbone=LLaMA-2-7B2025.1015.91
ELSAModel=Gemma-2-2B, Sparsity=60%2025.1015.93
VanillaModel Size=Pythia-1.4B, Scheme=Vanilla, Layers=242025.1015.97
AWQ (W4A16)Model=Gemma-2-9B, Calibration Method=COLA2026.0316.01
AWQ (W4A16)Model=Gemma-2-9B, Calibration Method=ZipCal2026.0316.01
AWQ (W4A16)Model=Gemma-2-9B, Calibration Method=ZipCal (Multi-Domain)2026.0316.01
MagnitudeModel=LLaMA-2-7B, Sparsity=50%2025.1016.03
SAFEModel=OPT-1.3B, Sparsity=50%2025.1016.38
SparseGPTModel=Gemma-2-2B, Sparsity=60%2025.1016.53
ELSAModel=LLaMA-3.2-3B, Sparsity=60%2025.1016.65