Humanity's Last Exam (HLE)
Expert-written closed questions at the frontier of knowledge, multimodal
knowledge and reasoningcrowded
Score in %
Definition
- Version
- not in the notes
- Category
- knowledge and reasoning
- Maintainer
- center-for-ai-safety and scale-ai
- Status
- crowded
- Home page
- https://arxiv.org/abs/2501.14249
Sources
- huggingface.co/moonshotai/Kimi-K3 https://huggingface.co/moonshotai/Kimi-K3
- Humanity's Last Exam arxiv.org https://arxiv.org/abs/2501.14249
- HLE Leaderboard: Claude Fable 5.1 Leads at 65% benchlm.ai https://www.benchlm.ai/benchmarks/hle
Scores at a Glance
Score in %
Higher is better for this unit.
1
GLM-5.3Z.ai (Zhipu)with tools
Kimi K3Moonshot AIwith tools
Step-3.7-FlashStepFunwith tools
Qwen3.8-2.4T-A95BAlibaba (Qwen)
DeepSeek-V4-ProDeepSeek
GLM-5.3Z.ai (Zhipu)with tools62.5 %
vendor-reported2Kimi K3Moonshot AIwith tools56 %
vendor-reported3Step-3.7-FlashStepFunwith tools48.1 %
vendor-reported4Qwen3.8-2.4T-A95BAlibaba (Qwen)43.6 %
vendor-reported5DeepSeek-V4-ProDeepSeek37.7 %
vendor-reported0255075100
Where this comes from
Each bar is one model, with its most independent recorded score. Hatched bars are vendor-reported; the word beside each bar says who ran it. Settings and dates differ between rows; the Scores table below has every result.
Scores
| Model | Score | Who ran it | Setting | View or population | Date | Source |
|---|---|---|---|---|---|---|
 Claude Fable 5.1 | 65 % | vendor-reported | Table mixes runs with and without tools | rank 1, BenchLM aggregator, top score | 10 Oct 2026 | benchlm.ai |
| GLM-5.3 | 62.5 % | vendor-reported | with tools | – | – | huggingface.co |
| Kimi K3 | 56 % | vendor-reported | with tools | – | – | huggingface.co |
 Gemini 3 Deep Think | 48.4 % | vendor-reported | Without tools; updated Deep Think | Gemini 3 Deep Think upgrade post | 12 Feb 2026 | blog.google |
| Step-3.7-Flash | 48.1 % | vendor-reported | with tools | – | – | huggingface.co |
| Qwen3.8-2.4T-A95B | 43.6 % | vendor-reported | – | – | – | huggingface.co |
| Kimi K3 | 43.5 % | vendor-reported | without tools | – | – | huggingface.co |
| Gemini 3 Deep Think | 41 % | vendor-reported | Without tools | Gemini 3 launch post | 18 Nov 2025 | blog.google |
| DeepSeek-V4-Pro | 37.7 % | vendor-reported | – | – | – | huggingface.co |
 Motif-3 | 37 % | vendor-reported | temperature 1.0, top-p 0.95, max sequence 262,144 | Motif-3 model card | Aug 2026 | huggingface.co |
 A.X K2 | 27.8 % | vendor-reported | Thinking mode | A.X K2 model card | Jul 2026 | huggingface.co |
 Ling 3.0 Flash (Ling-3.0-flash) | 22.7 % | vendor-reported | – | Ling-3.0-flash model card eval results | Jul 2026 | huggingface.co |
 MiMo-V2-Flash | 22.1 % | vendor-reported | No tools | MiMo-V2-Flash model card, post-trained model table | 16 Dec 2025 | huggingface.co |
Where this comes from
Scores from different settings, harnesses or dates are not directly comparable; the row says which. Sorted by score, highest first, as a reading aid, not a ranking.
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Humanity's Last Exam (HLE)", https://aurion.semai.id/benchmarks/hle/, records of October 2026.