List your company

Long-context performance evaluation on RULER

96AccuracyFP16
9.6832.0954.576.91Feb 7, 2026Mar 3, 2026Mar 27, 2026Apr 20, 2026May 14, 2026Jun 7, 2026Jul 1, 2026

Evaluation results

MethodLinks
FP16Backbone=LLaMA-3-8B-Instruct, BPA=162026.0796——————
GSRQBackbone=LLaMA-3-8B-Instruct, BPA=22026.0796——————
DenseModel=Llama-3.1-8B-Instruct, Context Length=4K2026.0295.31——————
DenseModel=Llama-3.1-8B-Instruct, Context Length=8K2026.0295.31——————
Sketch&WalkModel=Llama-3.1-8B-Instruct, Context Length=8K2026.0295——————
DenseModel=Llama-3.1-8B-Instruct, Context Length=16K2026.0294.06——————
KIVIBackbone=LLaMA-3-8B-Instruct, BPA=22026.0794——————
VQLLMBackbone=LLaMA-3-8B-Instruct, BPA=22026.0794——————
Sketch&WalkModel=Llama-3.1-8B-Instruct, Context Length=4K2026.0293.64——————
Sketch&WalkModel=Llama-3.1-8B-Instruct, Context Length=16K2026.0293.13——————
DenseModel=Llama-3.1-8B-Instruct, Context Length=32K2026.0289.69——————
Sketch&WalkModel=Llama-3.1-8B-Instruct, Context Length=32K2026.0287.5——————
DenseModel=Llama-3.1-8B-Instruct, Context Length=64K2026.0282.81——————
Sketch&WalkModel=Llama-3.1-8B-Instruct, Context Length=64K2026.0281.5——————
GSRQBackbone=LLaMA-3-8B-Instruct, BPA=12026.0772——————
VQLLMBackbone=LLaMA-3-8B-Instruct, BPA=12026.0720——————
KIVIBackbone=LLaMA-3-8B-Instruct, BPA=12026.0713——————
DenseModel=Llama-3.1-8B-Instruct, Context Length=Average2026.02—91.44—————
FastKVPrefill=60%, KV=10%, Backbone=LLaMA-3.1-8B-Instruct2025.02—75.677.877.377.277.468.2
Full-contextPrefill=100%, KV=100%, Backbone=LLaMA-3.1-8B-Instruct2025.02—8690.19583.485.576.3
GemFilterPrefill=51%, KV=10%, Backbone=LLaMA-3.1-8B-Instruct2025.02—69.669.768.270.469.869.8
H2OPrefill=100%, KV=10%, Backbone=LLaMA-3.1-8B-Instruct2025.02——27.1————
PyramidInferPrefill=60%, KV=60%, Backbone=LLaMA-3.1-8B-Instruct2025.02——66.5————
Sketch&WalkModel=Llama-3.1-8B-Instruct, Context Length=Average2026.02—90.15—————
SnapKVPrefill=100%, KV=10%, Backbone=LLaMA-3.1-8B-Instruct2025.02—73.675.676.872.97567.7
StreamingLLMPrefill=100%, KV=10%, Backbone=LLaMA-3.1-8B-Instruct2025.02—18.61521.524.416.915