List your company

Language Modeling on OpenWebText2

5.45Perplexity (PPL)Full KV
4.22612.48820.7529.012May 6, 2026May 8, 2026May 11, 2026May 13, 2026May 16, 2026May 18, 2026May 21, 2026

Evaluation results

MethodLinks
Full KVcontext length=4k, Backbone=Llama-3.1-8B-Instruct, KV cache compression (B)=Full2026.055.45—
Meta-Softcontext length=4k, Backbone=Llama-3.1-8B-Instruct, KV cache compression (B)=2562026.055.68—
Judge Qcontext length=4k, Backbone=Llama-3.1-8B-Instruct, KV cache compression (B)=2562026.055.71—
Full KVcontext length=16k, Backbone=Llama-3.1-8B-Instruct, KV cache compression (B)=Full2026.055.72—
ZeroMergecontext length=4k, Backbone=Llama-3.1-8B-Instruct, KV cache compression (B)=2562026.055.75—
H2Ocontext length=4k, Backbone=Llama-3.1-8B-Instruct, KV cache compression (B)=2562026.055.82—
SnapKVcontext length=4k, Backbone=Llama-3.1-8B-Instruct, KV cache compression (B)=2562026.055.83—
Meta-Softcontext length=16k, Backbone=Llama-3.1-8B-Instruct, KV cache compression (B)=2562026.055.87—
Judge Qcontext length=16k, Backbone=Llama-3.1-8B-Instruct, KV cache compression (B)=2562026.055.94—
ZeroMergecontext length=16k, Backbone=Llama-3.1-8B-Instruct, KV cache compression (B)=2562026.055.98—
SnapKVcontext length=16k, Backbone=Llama-3.1-8B-Instruct, KV cache compression (B)=2562026.056.09—
H2Ocontext length=16k, Backbone=Llama-3.1-8B-Instruct, KV cache compression (B)=2562026.056.12—
AIR-MoESize=Large2026.0516.6511.2
Coarse MoESize=Large2026.0516.9811.3
PEERSize=Large2026.0517.8811.2
Hierarchical MoESize=Large2026.0518.1711.3
AIR-MoESize=Medium2026.0520.513.6
Coarse MoESize=Medium2026.0521.253.5
PEERSize=Medium2026.0521.33.5
Hierarchical MoESize=Medium2026.0523.073.5
Standard Granular MoESize=Small2026.0531.56648.3
AIR-MoESize=Small2026.0532.14505
PEERSize=Small2026.0533.1505.6
Hierarchical MoESize=Small2026.0536.05502.6
Coarse MoESize=Small2026.0536.05493.9