Benchmark
LongBench v2
Long context benchmarks measure how well models use information in long inputs. Scores are percentages. Higher is better.
- Models and systems measured
- 27
- Scores recorded
- 34
- Sources
- 2
- IQ weight
- 2.68
Score distribution
- 58.8%
- Mean
- 61%
- Median
- 8.8%
- Std dev
- 46.4%
- 10th percentile
- 66.9%
- 90th percentile
- 26.1%
- Min
This benchmark has a weight of 2.68 for chat models, measured over 26 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.
Leaderboard
| # | Model or system | Score |
|---|---|---|
| 1 | Claude Opus 4.8Qwen3.8 官方模型卡对照表· Reported by Qwen· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 69.1% |
| 2 | Beam 1.0 501B A23B· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 67.5% |
| 3 | GPT-5.6 SolQwen3.8 官方模型卡对照表· Reported by Qwen· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 67.1% |
| 4 | Inkling 2 Mini· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 66.9% |
| 5 | Qwen3.8 Max· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 66.3% |
| 6 | Qwen3.8 MaxQwen3.8 官方模型卡对照表· Reported by Qwen· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 66.3% |
| 7 | Qwen3.7 MaxQwen3.8 官方模型卡对照表· Reported by Qwen· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 65.3% |
| 8 | Qwen3.5 397B A17B· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 63.2% |
| 9 | Qwen3.5 397B A17BQwen3.5 release/model card · published setting· Reported by Qwen· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 63.2% |
| 10 | Qwen3.6 Plus· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 62% |
| 11 | Qwen3.6 PlusQwen official release table · thinking mode unless benchmark requires otherwise· Reported by Alibaba Qwen· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 62% |
| 12 | Step 3.5 FlashStep 3.5 technical report · Table 5 · short + medium subsets · PaCoRe· Reported by StepFun· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 62% |
| 13 | Nemotron 3 Ultra· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 61.9% |
| 14 | Nemotron 3 UltraNeMo Evaluator SDK · ≤1M· Reported by NVIDIA· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 61.9% |
| 15 | MiniMax M1 80K· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. | 61.5% |
| 16 | Kimi K2.5setting not named by the source· Reported by Moonshot AI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 61% |
| 17 | Kimi K2.5· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 61% |
| 18 | Mai Thinking 1· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 61% |
| 19 | MiniMax M1 40K· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. | 61% |
| 20 | MiMo V2 Flash· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 60.6% |
| 21 | MiMo V2 FlashMiMo-V2-Flash final evaluation · published post-training table· Reported by Xiaomi MiMo· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 60.6% |
| 22 | Qwen3 Max· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 60.6% |
| 23 | Qwen3.5 27B· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 60.6% |
| 24 | Qwen3.5 122B-A10B· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 60.2% |
| 25 | Qwen3.5 Omni PlusQwen3.5-Omni release · no-thinking setting· Reported by Qwen, Qwen Team· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 59.6% |
| 26 | Qwen3.5 35B A3B· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 59% |
| 27 | Step 3.5 FlashStep 3.5 technical report · Table 5 · short + medium subsets · vanilla· Reported by StepFun· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 57.5% |
| 28 | Qwen3.5 9B· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 55.2% |
| 29 | Qwen3.5 4B· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 50% |
| 30 | DeepSeek V3· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 48.7% |
| 31 | Qwen3.5 Omni FlashQwen3.5-Omni release · no-thinking setting· Reported by Qwen, Qwen Team· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 46.4% |
| 32 | Qwen3.5 2B· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 38.7% |
| 33 | Qwen3.5 0.8B· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 26.1% |
Not ranked
The source marks these results as descriptive, not capability scores, so we list them by name instead of ranking them.
| Model or system | Score |
|---|---|
| DeepSeek V4.1 Flash BaseEM · 1-shot· Reported by DeepSeek· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 45.2% |
Data credits: Data by LLM Stats · BenchBoard by SYuan03 (github.com/SYuan03/benchboard). MIT covers its code and data organization; no results licence is inferred. Numeric relays retain the original stating-source links.