List your company

Long-context Question Answering on ∞Bench

78.46AccuracyStateLM-14B-RL
21.842436.541251.2465.9388Feb 12, 2026

Evaluation results

MethodLinks
StateLM-14B-RLContext=32K2026.0278.46——
StateLM-14BContext=32K2026.0277.44——
Qwen3-14BContext=128K2026.0274.96——
RL-MemoryAgent-14BContext=32K2026.0274.24——
Qwen3-235B (w/ Pensieve)Context=256K2026.0273.36——
StateLM-8B-RLContext=32K2026.0273.07——
StateLM-8BContext=32K2026.0270.16——
StateLM-4BContext=32K2026.0267.25——
Qwen3-8BContext=128K2026.0266.81——
RL-MemoryAgent-7BContext=32K2026.0262.45——
Qwen3-4BContext=128K2026.0259.97——
ReadAgent-14BContext=32K2026.0234.06——
ReadAgent-8BContext=32K2026.0224.02——
Qwen3-0.6BAttention=Full attention, Avg. Len.=128k2026.02—1.762.18
Qwen3-0.6B-ALLMEMArchitecture=ALLMEM, Avg. Len.=128k2026.02—8.5312.54
Qwen3-0.6B-MambaArchitecture=Hybrid Mamba-2, Avg. Len.=128k2026.02—68.5
Qwen3-0.6B-SWA-SinksAttention=sliding-window, Window Length=8k, Attention Sinks=128, Avg. Len.=128k2026.02—3.74.05