List your company

Long-context language modeling evaluation on RULER

100Single-key AccuracyFull Attention
41.7656.887287.12Feb 6, 2025Apr 13, 2025Jun 18, 2025Aug 23, 2025Oct 28, 2025Jan 2, 2026Mar 9, 2026

Evaluation results

MethodLinks
Full AttentionContext Length=4k2026.0310010010010082.584.489.859.289.5——
LycheeClusterContext Length=4k2026.0310010010099.581.280.389.859.288.8——
LycheeClusterContext Length=8k2026.0310010096.41007883.781.659.287.4——
Full CacheBackbone=Llama-3.1-8B, Cache Size=100%, Context Length=32K2025.0210010099.7598.2599.495.33846291.0544.90
SnapKVBackbone=Llama-3.1-8B, Cache Size=40%, Context Length=32K2025.021009387.592.7596.690415267.9320.325.4
SnapKV + CriticalKVBackbone=Llama-3.1-8B, Cache Size=40%, Context Length=32K2025.0210010099999792.67635376.8933.915.6
AdaKV + CriticalKVBackbone=Llama-3.1-8B, Cache Size=40%, Context Length=32K2025.0210010098.7599.2599.694.67645886.2852.45.2
HeadKVBackbone=Llama-3.1-8B, Cache Size=40%, Context Length=32K2025.021009993979892.67565479.9817.112.2
HeadKV + CriticalKVBackbone=Llama-3.1-8B, Cache Size=40%, Context Length=32K2025.021001009999.599.493.33725989.2956.51.9
Full CacheBackbone=Qwen2.5-32B, Cache Size=100%, Context Length=32K2025.021009999.2510010096827493.9190.60
SnapKVBackbone=Qwen2.5-32B, Cache Size=40%, Context Length=32K2025.021006361.7571.7599.892506463.8687.932
SnapKV + CriticalKVBackbone=Qwen2.5-32B, Cache Size=40%, Context Length=32K2025.021009899.599.7510093.33676981.0988.613.7
AdaKVBackbone=Qwen2.5-32B, Cache Size=40%, Context Length=32K2025.021008183.587.510093.33556471.0988.924.3
AdaKV + CriticalKVBackbone=Qwen2.5-32B, Cache Size=40%, Context Length=32K2025.021009899.510010094.33646683.8789.510.7
HeadKVBackbone=Qwen2.5-32B, Cache Size=40%, Context Length=32K2025.021009296.2597.2510094.67686681.0489.313.7
HeadKV + CriticalKVBackbone=Qwen2.5-32B, Cache Size=40%, Context Length=32K2025.021009998.7510010095.33726990.6989.93.4
AdaKVBackbone=Llama-3.1-8B, Cache Size=40%, Context Length=32K2025.02999898.2597.598.692435478.3826.613.9
Full CacheBackbone=Mistral-7B, Cache Size=100%, Context Length=32K2025.0299949594.2590.695.67605978.27300
AdaKV + CriticalKVBackbone=Mistral-7B, Cache Size=40%, Context Length=32K2025.029989958589.495.33605769.1740.511.6
Full AttentionContext Length=32k2026.0395.995.997.597.576.376.979.659.284.8——
LycheeClusterContext Length=32k2026.0395.99892.993.985.776.275.559.284.7——
LycheeClusterContext Length=16k2026.0393.910094.994.973.189.877.65184.4——
Full AttentionContext Length=8k2026.0391.81001009979.687.181.659.287.3——
HeadKV + CriticalKVBackbone=Mistral-7B, Cache Size=40%, Context Length=32K2025.02845362.2560.581.495535557.5949.526.4
SnapKV + CriticalKVBackbone=Mistral-7B, Cache Size=40%, Context Length=32K2025.02743124.252681.894.67455341.5648.546.9
HeadKVBackbone=Mistral-7B, Cache Size=40%, Context Length=32K2025.02702416.516.2582.692.67455339.5930.649.4
SnapKVBackbone=Mistral-7B, Cache Size=40%, Context Length=32K2025.026516811.2574.891.67345132.1532.258.9
Full AttentionContext Length=16k2026.0363.310098.598.58293.979.65183.3——
AdaKVBackbone=Mistral-7B, Cache Size=40%, Context Length=32K2025.02442514.252071.692.33445534.8825.355.4