Skip to content

LongBench v2

A long-context benchmark that measures whether models can actually use extended context windows for reasoning and retrieval.

17 models

Higher is better.

  1. 1Beam 1.0 501B-A23BReflection AI67.5
  2. 2Qwen3.8 MaxAlibaba66.3
  3. 3Claude Opus 4.5Anthropic64.4
  4. 4Qwen3.5 397BAlibaba63.2
  5. 5Qwen3.6 PlusAlibaba62.0
  6. 6Nemotron 3 UltraNVIDIA61.9
  7. 7MiniMax M1 80kMiniMax61.5
  8. 8Kimi K2.5 (Reasoning)Moonshot AI61.0
  9. 9Kimi K2.5Moonshot AI61.0
  10. 10GLM-5Z.AI60.8
  11. 11Qwen3.5-27BAlibaba60.6
  12. 12MiMo-V2-FlashXiaomi60.6
  13. 13Agents-A1InternScience60.2
  14. 14Qwen3.5-122B-A10BAlibaba60.2
  15. 15Qwen3.5-35B-A3BAlibaba59.0
  16. 16DeepSeek V3DeepSeek48.7
  17. 17MiniCPM5-2BOpenBMB43.7
Higher is better. Bar colour: blue proprietary, orange open weights, green other.

About this benchmark

Metric
Extended-context retrieval and reasoning
Tasks
Long-context tasks
Authors
LongBench v2 authors (2025)

Scores are as reported by model developers or evaluators and may use different settings (for example reasoning effort or tools). Compare with care.