// BENCHMARK

DROP benchmark

AI model leaderboard for the DROP benchmark. Compare how large language models score on DROP, see the full ranking, and understand what this AI benchmark measures. Llama 3.1-405B currently leads with 92.5%. Discrete reasoning over paragraphs — reading comprehension with arithmetic.

Leaderboard

# Model Organization Score Variant Source
#1 Llama 3.1-405B Meta AI 92.5% instruct-f1-cited-minimax-text official ↗
#2 DeepSeek-V3 DeepSeek 91.8% F1 official ↗
#3 Pangu Pro MoE Huawei 91.2% instruct-f1 official ↗
#4 DeepSeek-V3 DeepSeek 91% f1-cited-minimax-text official ↗
#5 Llama 4 Scout Meta AI 89.3% f1-cited-pangu-pro-moe official ↗
#6 Gemini 2.0 Flash Google DeepMind 89.3% exp-f1-cited-minimax-text official ↗
#7 GPT-4o (Nov 2024) OpenAI 89.2% f1-cited-minimax-text official ↗
#8 Gemini 1.5 Pro Google DeepMind 89.2% 002-f1-cited-minimax-text official ↗
#9 Claude 3.5 Sonnet Anthropic 88.8% 1022-f1-cited-minimax-text official ↗
#10 Claude 3.5 Sonnet Anthropic 88.3% 1022-cited-mimo official ↗
#11 MiniMax-Text-01 MiniMax 87.8% f1-from-hf-readme official ↗
#12 dots.llm1 Rednote 87% F1 official ↗
#13 DeepSeek-R1-Distill-Qwen-14B DeepSeek 85.5% cited-mimo official ↗
#14 DeepSeek-R1-Distill-Qwen-14B Qwen 85.5% cited-mimo official ↗
#15 Qwen2.5 Instruct (72B) Qwen 85% instruct-f1-cited-minimax-text official ↗
#16 Llama 3.1-70B Meta AI 77% instruct-cited-tulu3 official ↗
#17 DeepSeek-R1-Distill-Qwen-7B Qwen 77% cited-mimo official ↗
#18 DeepSeek-R1-Distill-Qwen-7B DeepSeek 77% cited-mimo official ↗
#19 OLMo 2 32B Allen Institute for AI 73.6% base-from-hf-readme official ↗
#20 Hermes 3 70B Nous Research 73.2% cited-tulu3 official ↗
#21 Gemma 2 9B Google DeepMind 63% base-cited-olmo2 official ↗
#22 Tulu 3 8B Allen Institute for AI 62.6% instruct-3-shot official ↗
#23 Llama 3.1-8B Meta AI 61.5% instruct-cited-tulu3 official ↗
#24 Llama 3.1-8B Meta AI 61.48% instruct-cited-granite-3-2 official ↗
#25 Gemma 2 9B Google DeepMind 58.8% instruct-cited-tulu3 official ↗
#26 Llama 3.1-8B Meta AI 56.4% base-cited-olmo2 official ↗
#27 Ministral 8B Mistral AI 56.2% instruct-cited-tulu3 official ↗
#28 Qwen2.5-7B Qwen 55.8% base-cited-olmo2 official ↗
#29 Qwen2.5-7B Qwen 54.71% instruct-cited-granite-3-2 official ↗
#30 Granite 3.2 8B IBM 50.95% instruct-from-hf-readme official ↗
#31 Granite 3.1 8B IBM 50.78% instruct-cited-granite-3-2 official ↗
#32 Llama 2-13B Meta AI 45.6% base-cited-olmo2 official ↗
#33 DeepSeek-R1-Distill-Llama-8B DeepSeek 44.46% cited-granite-3-2 official ↗
#34 DeepSeek-R1-Distill-Qwen-7B DeepSeek 42.76% cited-granite-3-2 official ↗
#35 DeepSeek-R1-Distill-Qwen-7B Qwen 42.76% cited-granite-3-2 official ↗
#36 Qwen2.5 Instruct (72B) Qwen 34.2% instruct-cited-tulu3 official ↗
#37 Granite 3.2 2B IBM 21.12% instruct-from-hf-readme official ↗
#38 Granite 3.1 2B IBM 18.68% instruct-cited-granite-3-2 official ↗

Frequently asked questions about DROP

What is the DROP benchmark?

Discrete reasoning over paragraphs — reading comprehension with arithmetic.

How is the DROP benchmark scored?

DROP is scored using the F1 (%) metric, where a higher score is better. The maximum achievable score is 100.000. GenAIList aggregates reported scores from model providers and papers into a single ranked leaderboard.

Which AI model scores highest on DROP?

As of the latest reported scores on GenAIList, Llama 3.1-405B achieves the highest result on DROP with a score of 92.5%.

Is a higher DROP score better?

Yes. On DROP a higher score indicates better performance, so models near the top of the leaderboard are the strongest.