// BENCHMARK

IFEval benchmark

AI model leaderboard for the IFEval benchmark. Compare how large language models score on IFEval, see the full ranking, and understand what this AI benchmark measures. GPT-5 currently leads with 95.38%. Verifiable instruction-following benchmark with 25 instruction types.

Leaderboard

# Model Organization Score Variant Source
#1 GPT-5 OpenAI 95.38% thinking-high-cited-ring-1t official ↗
#2 Grok 4 xAI 92.4% official ↗
#3 LFM2.5-8B-A1B Liquid AI 91.84% official ↗
#4 Kimi K2 Moonshot 90.99% instruct-0905-prompt-strict-cited-ling-1t official ↗
#5 gemini-2.5-pro Google DeepMind 90.8% official ↗
#6 NVIDIA-Nemotron-Nano-9B-v2 NVIDIA 90.3% from-hf-readme official ↗
#7 Claude 3.5 Sonnet Anthropic 90.1% 1022-avg-cited-minimax-text official ↗
#8 Granite-4.0-H-Small IBM 89.87% instruct-strict official ↗
#9 Kimi K2 Moonshot 89.8% official ↗
#10 LongCat-Flash Meituan Inc 89.65% official ↗
#11 Gemini 1.5 Pro Google DeepMind 89.4% 002-avg-cited-minimax-text official ↗
#12 MiniMax-Text-01 MiniMax 89.1% avg-from-hf-readme official ↗
#13 DeepSeek-V3.1-Terminus DeepSeek 89.09% thinking-cited-ring-1t official ↗
#14 gemini-2.5-pro Google DeepMind 88.72% cited-ring-1t official ↗
#15 Claude Sonnet 4 Anthropic 88.7% official ↗
#16 Gemini 2.0 Flash Google DeepMind 88.4% exp-avg-cited-minimax-text official ↗
#17 Llama 3.1-405B Meta AI 88.4% instruct-cited-tulu3 official ↗
#18 DeepSeek-V3 DeepSeek 88% cited-tulu3 official ↗
#19 Llama 3.1-70B Meta AI 88% instruct-cited-tulu3 official ↗
#20 Qwen3-235B-A22B-Thinking (Jul 2025) Qwen 87.8% thinking-cited-ring-1t official ↗
#21 DeepSeek-R1-0528 DeepSeek 87.8% official ↗
#22 Qwen2.5 Instruct (72B) Qwen 87.6% instruct-cited-tulu3 official ↗
#23 Granite-4.0-H-Small IBM 87.55% average official ↗
#24 Llama 3.1-70B Meta AI 87.5% cited-telechat2 official ↗
#25 GPT-4.1 OpenAI 87.4% official ↗
#26 DeepSeek-V3 DeepSeek 87.3% avg-cited-minimax-text official ↗
#27 Qwen2.5 Instruct (72B) Qwen 87.2% instruct-avg-cited-minimax-text official ↗
#28 gemini-2.5-pro Google DeepMind 87.08% lowthink-prompt-strict-cited-ling-1t official ↗
#29 IBM Granite 4.1 IBM 87.06%
#30 Granite-4.0-H-Micro IBM 86.94% instruct-strict official ↗
#31 Claude 3.5 Sonnet Anthropic 86.5% 1022-cited-mimo official ↗
#32 Qwen3-30B-A3B Qwen 86.5% thinking-strict-prompt official ↗
#33 Llama 3.1-405B Meta AI 86.4% instruct-avg-cited-minimax-text official ↗
#34 DeepSeek-V3.1-Terminus DeepSeek 86.32% non-thinking-prompt-strict-cited-ling-1t official ↗
#35 GLM-4.5-Air Zhipu AI 86.3% official ↗
#36 Ling-1T Ant Group 86.11% prompt-strict official ↗
#37 GLM 4.5 Zhipu AI 86.1% official ↗
#38 DeepSeek-V3 DeepSeek 86.1% official ↗
#39 Tulu 3 405B Allen Institute for AI 86% instruct-prompt-loose official ↗
#40 Pangu Pro MoE Huawei 85.7% instruct-prompt-strict official ↗
#41 Qwen3-14B Qwen 85.4% thinking-strict-prompt official ↗
#42 Llama 4 Maverick Meta AI 85.4% official ↗
#43 Granite-4.0-H-Small IBM 85.22% prompt-strict official ↗
#44 Ring-1T Ant Group 85.21% thinking official ↗
#45 Llama 4 Scout Meta AI 85.2% prompt-strict-cited-pangu-pro-moe official ↗
#46 GPT-5 OpenAI 85.11% main-prompt-strict-cited-ling-1t official ↗
#47 Qwen3-32B Qwen 85% reasoning official ↗
#48 Qwen3-32B Qwen 85% thinking-strict-prompt official ↗
#49 Qwen3-8B Qwen 85% thinking-strict-prompt official ↗
#50 Phi-4-Reasoning-plus Microsoft 84.9% reasoning official ↗
#51 GPT-4o (Nov 2024) OpenAI 84.8% cited-tulu3 official ↗
#52 Qwen3-14B Qwen 84.8% non-thinking-strict-prompt official ↗
#53 Granite-4.0-H-Tiny IBM 84.78% instruct-strict official ↗
#54 Granite-4.0-H-Micro IBM 84.32% average official ↗
#55 GPT-4o (Nov 2024) OpenAI 84.1% avg-cited-minimax-text official ↗
#56 Qwen3-30B-A3B Qwen 83.7% non-thinking-strict-prompt official ↗
#57 EXAONE 4.0 (32B) LG AI Research 83.7% reasoning official ↗
#58 Qwen3-235B-A22B-Thinking (Jul 2025) Qwen 83.4% reasoning official ↗
#59 Qwen3-235B-A22B Qwen 83.4% thinking-strict-prompt official ↗
#60 DeepSeek-V3-0324 DeepSeek 83.4% official ↗
#61 Qwen3-235B-A22B Qwen 83.2% non-thinking-strict-prompt official ↗
#62 Qwen3-32B Qwen 83.2% non-thinking-strict-prompt official ↗
#63 Qwen3-8B Qwen 83% non-thinking-strict-prompt official ↗
#64 Telechat2-115B China Telecom 82.81% from-hf-readme official ↗
#65 Gemma 3 27B Google DeepMind 82.6% official ↗
#66 Tulu 3 8B Allen Institute for AI 82.4% instruct-prompt-loose official ↗
#67 dots.llm1 Rednote 82.1% prompt strict official ↗
#68 Hermes 3 405B Nous Research 81.9% cited-tulu3 official ↗
#69 Qwen3-4B Qwen 81.9% thinking-strict-prompt official ↗
#70 Granite-4.0-H-Micro IBM 81.71% prompt-strict official ↗
#71 EXAONE 3.5 32B LG AI Research 81.7% instruct-from-hf-readme official ↗
#72 Granite-4.0-H-Tiny IBM 81.44% average official ↗
#73 DeepSeek-V3-0324 DeepSeek 81.2% official ↗
#74 Qwen3-4B Qwen 81.2% non-thinking-strict-prompt official ↗
#75 DeepSeek-R1-0528 DeepSeek 80.8% reasoning official ↗
#76 Llama 3.1-8B Meta AI 80.6% instruct-cited-tulu3 official ↗
#77 Gemma 3 1B Google DeepMind 80.2% non-reasoning official ↗
#78 TeleChat2-35B China Telecom 79.63% from-hf-readme official ↗
#79 Llama 3.1-8B Meta AI 79.1% instruct-cited-granite-3-2 official ↗
#80 EXAONE 3.5 7.8B LG AI Research 78.9% instruct-from-hf-readme official ↗
#81 Llama 3.1-8B Meta AI 78.8% instruct-cited-falcon3 official ↗
#82 Qwen2.5-32B Qwen 78.7% instruct-cited-exaone35 official ↗
#83 DeepSeek-R1-Distill-Qwen-14B Qwen 78.3% cited-mimo official ↗
#84 DeepSeek-R1-Distill-Qwen-14B DeepSeek 78.3% cited-mimo official ↗
#85 Granite-4.0-H-Tiny IBM 78.1% prompt-strict official ↗
#86 Qwen2-72B Qwen 77.6% instruct-cited-telechat2 official ↗
#87 Falcon3-7B Technology Innovation Institute 76.5% instruct-from-hf-readme official ↗
#88 Hermes 3 70B Nous Research 76% cited-tulu3 official ↗
#89 Qwen2.5-7B Qwen 74.9% instruct-cited-granite-3-2 official ↗
#90 Qwen2.5-7B Qwen 74.7% instruct-cited-falcon3 official ↗
#91 Granite 3.2 8B IBM 74.31% instruct-from-hf-readme official ↗
#92 EXAONE 3.5 2.4B LG AI Research 73.6% instruct-from-hf-readme official ↗
#93 Granite 3.1 8B IBM 73.2% instruct-cited-granite-3-2 official ↗
#94 TeleChat2-7B China Telecom 73.1% from-hf-readme official ↗
#95 Qwen3-1.7B Qwen 72.5% thinking-strict-prompt official ↗
#96 Llama 3.2 3B Meta AI 70.1% instruct-cited-exaone35 official ↗
#97 Gemma 2 9B Google DeepMind 69.9% instruct-cited-tulu3 official ↗
#98 Qwen3-1.7B Qwen 68.2% non-thinking-strict-prompt official ↗
#99 Qwen3-1.7B Qwen 68.2% non-reasoning official ↗
#100 EXAONE 4.0 (1.2B) LG AI Research 67.8% reasoning official ↗
#101 DeepSeek-R1-Distill-Llama-8B DeepSeek 66.5% cited-granite-3-2 official ↗
#102 DeepSeek-V2 (MoE-236B) DeepSeek 63.8% cited-telechat2 official ↗
#103 Granite 3.1 2B IBM 63.59% instruct-cited-granite-3-2 official ↗
#104 Granite 3.2 2B IBM 61.55% instruct-from-hf-readme official ↗
#105 TeleChat2-3B China Telecom 61.29% from-hf-readme official ↗
#106 Qwen2.5-3B Qwen 60.8% instruct-cited-exaone35 official ↗
#107 DeepSeek-R1-Distill-Qwen-7B DeepSeek 60.5% cited-mimo official ↗
#108 DeepSeek-R1-Distill-Qwen-7B Qwen 60.5% cited-mimo official ↗
#109 Gemma 2 27B Google DeepMind 59.7% instruct-cited-exaone35 official ↗
#110 Qwen3-0.6B Qwen 59.2% thinking-strict-prompt official ↗
#111 DeepSeek-R1-Distill-Qwen-7B DeepSeek 59.1% cited-granite-3-2 official ↗
#112 DeepSeek-R1-Distill-Qwen-7B Qwen 59.1% cited-granite-3-2 official ↗
#113 Ministral 8B Mistral AI 56.4% instruct-cited-tulu3 official ↗
#114 Yi-1.5-34B 01.AI 55.5% instruct-cited-exaone35 official ↗
#115 Qwen3-0.6B Qwen 54.5% non-reasoning official ↗
#116 Qwen3-0.6B Qwen 54.5% non-thinking-strict-prompt official ↗
#117 Gemma 2 2B Google DeepMind 50.5% instruct-cited-exaone35 official ↗
#118 Qwen2.5-1.5B Qwen 40.7% instruct-cited-exaone35 official ↗
#119 Falcon Mamba Technology Innovation Institute 33.36% 0-shot-from-hf-readme official ↗
#120 Gemma 7B Google DeepMind 26.59% cited-falcon-mamba official ↗
#121 Mistral 7B Mistral AI 23.86% v0-1-cited-falcon-mamba official ↗
#122 Llama 3-8B Meta AI 14.55% cited-falcon-mamba official ↗
#123 Llama 3.1-8B Meta AI 12.7% base-cited-falcon-mamba official ↗

Frequently asked questions about IFEval

What is the IFEval benchmark?

Verifiable instruction-following benchmark with 25 instruction types.

How is the IFEval benchmark scored?

IFEval is scored using the accuracy (%) 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 IFEval?

As of the latest reported scores on GenAIList, GPT-5 achieves the highest result on IFEval with a score of 95.38%.

Is a higher IFEval score better?

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