LongBench v2
A long-context benchmark that measures whether models can actually use extended context windows for reasoning and retrieval.
Top models on LongBench v2 — November 21, 2026
As of November 21, 2026, Inkling 2 Small leads the LongBench v2 leaderboard with 70.6% , followed by Beam 1.0 501B-A23B FP8 (67.6%) and Beam 1.0 501B-A23B (67.5%).
Inkling 2 Small
Thinking Machines Lab
Reported by Thinking Machines LabBeam 1.0 501B-A23B FP8
Reflection AI
Reported by Reflection AIBeam 1.0 501B-A23B
Reflection AI
Reported by Reflection AI27 modelsReasoning6% of Reasoning reference weightCurrentUpdated November 21, 2026
Leaderboard (27 models)
None of the 27 results on this page has a documented run by the benchmark or an independent evaluator: 25 reported by the labs that built the models and 2 reported in another publisher’s document.
| Rank | Model / configuration | Score | Who ran it | Parameters (B) | Open / closed |
|---|---|---|---|---|---|
| 1 | Inkling 2 SmallThinking Machines LabReported by Thinking Machines Lab | 70.6% | Reported by Thinking Machines Lab | Not reported | Closed |
| 2 | Beam 1.0 501B-A23B FP8Reflection AIReported by Reflection AI | 67.6% | Reported by Reflection AI | Not reported | Open |
| 3 | Beam 1.0 501B-A23BReflection AIReported by Reflection AI | 67.5% | Reported by Reflection AI | Not reported | Open |
| 4 | Beam 1.0 501B-A23B NVFP4Reflection AIReported by Reflection AI | 67.2% | Reported by Reflection AI | Not reported | Open |
| 5 | Inkling 2 MiniThinking Machines LabReported by Thinking Machines Lab | 66.9% | Reported by Thinking Machines Lab | Not reported | Open |
| 6 | Qwen3.8 MaxAlibabaReported by Alibaba | 66.3% | Reported by Alibaba | Not reported | Open |
| 7 | Claude Opus 4.5AnthropicAnother lab’s comparison table | 64.4% | Another lab’s comparison table | Not reported | Closed |
| 8 | Qwen3.5 397BAlibabaReported by Alibaba | 63.2% | Reported by Alibaba | Not reported | Open |
| 9 | Qwen3.6 PlusAlibabaReported by Alibaba | 62% | Reported by Alibaba | Not reported | Closed |
| 10 | Nemotron 3 UltraNVIDIAReported by NVIDIA | 61.9% | Reported by NVIDIA | Not reported | Open |
| 11 | MiniMax M1 80kMiniMaxReported by MiniMax | 61.5% | Reported by MiniMax | Not reported | Closed |
| 12 | Kimi K2.5Moonshot AIReported by Moonshot AI | 61% | Reported by Moonshot AI | Not reported | Open |
| 13 | GLM-5Z.AIAnother lab’s comparison table | 60.8% | Another lab’s comparison table | Not reported | Open |
| 14 | Qwen3.5-27BAlibabaReported by Alibaba | 60.6% | Reported by Alibaba | Not reported | Open |
| 15 | MiMo-V2-FlashXiaomiReported by Xiaomi | 60.6% | Reported by Xiaomi | Not reported | Open |
| 16 | Qwen3.5-122B-A10BAlibabaReported by Alibaba | 60.2% | Reported by Alibaba | Not reported | Open |
| 17 | Agents-A1InternScienceReported by InternScience | 60.2% | Reported by InternScience | Not reported | Open |
| 18 | Qwen3.5-35B-A3BAlibabaReported by Alibaba | 59% | Reported by Alibaba | Not reported | Open |
| 19 | Spark-X2.6-4BSparkLLMReported by SparkLLM | 53.3% | Reported by SparkLLM | Not reported | Open |
| 20 | Agents-A1-4BInternScienceReported by InternScience | 52.1% | Reported by InternScience | Not reported | Open |
| 21 | DeepSeek V4 Pro BaseDeepSeekReported by DeepSeek | 51.5% | Reported by DeepSeek | Not reported | Open |
| 22 | Nemotron 3.5 Bolt E4B BF16NVIDIAReported by NVIDIA | 50.5% | Reported by NVIDIA | Not reported | Open |
| 23 | DeepSeek V3DeepSeekReported by DeepSeek | 48.7% | Reported by DeepSeek | Not reported | Open |
| 24 | Nemotron 3.5 Bolt E4BNVIDIAReported by NVIDIA | 48.7% | Reported by NVIDIA | Not reported | Open |
| 25 | DeepSeek V4 Flash BaseDeepSeekReported by DeepSeek | 44.7% | Reported by DeepSeek | Not reported | Open |
| 26 | MiniCPM5-2BOpenBMBReported by OpenBMB | 43.7% | Reported by OpenBMB | Not reported | Open |
| 27 | LLaDA2.2-miniInclusionAIReported by InclusionAI | 35.0% | Reported by InclusionAI | Not reported | Open |
- Lab-reported
- Benchmark-run
- Independent
According to BenchLM.ai, Inkling 2 Small leads the LongBench v2 benchmark with a score of 70.6%, followed by Beam 1.0 501B-A23B FP8 (67.6%) and Beam 1.0 501B-A23B (67.5%). The scores show moderate spread, with meaningful differences between the top tier and mid-tier models.
27 models have been evaluated on LongBench v2. The benchmark falls in the Reasoning category. BenchAlign v6.0 gives LongBench v2 6% of the Reasoning reference weight, so it moves the Reasoning leaderboard and the overall ranking. Reference weights are relative weights in the calibrated model, not fixed shares of a score.
About LongBench v2
Year
2025
Tasks
Long-context tasks
Format
Extended-context retrieval and reasoning
Difficulty
Hard long-context
LongBench v2 is useful because context-window size alone is not a capability. It measures whether a model can retain, retrieve, and reason over long inputs effectively.
Freshness and provenance
Version
LongBench v2 2025
Refresh cadence
Quarterly
Staleness state
Current
Question availability
Public benchmark set
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
What does LongBench v2 measure?
A long-context benchmark that measures whether models can actually use extended context windows for reasoning and retrieval.
Which model scores highest on LongBench v2?
Inkling 2 Small by Thinking Machines Lab currently leads with a score of 70.6% on LongBench v2.
How many models are evaluated on LongBench v2?
27 AI models have published results on LongBench v2 in the BenchLM catalog.
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 7,000+ subscribers.
One email each week. Unsubscribe anytime.