New: connect Claude & other AIs to GenAIList over MCP — research the catalog and contribute to the shared knowledge base. Learn how →
TruthfulQA benchmark
AI model leaderboard for the TruthfulQA benchmark. Compare how large language models score on TruthfulQA, see the full ranking, and understand what this AI benchmark measures. GPT-4o mini currently leads with 78.2%. 817 questions designed to elicit common human misconceptions and test factual honesty.
Leaderboard
| # | Model | Organization | Score | Variant | Source |
|---|---|---|---|---|---|
| #1 | GPT-4o mini | OpenAI | 78.2% | cited-phi35moe | official ↗ |
| #2 | Phi-3.5-MoE | Microsoft | 77.5% | instruct-mc2-10-shot | official ↗ |
| #3 | Gemini 1.5 Flash (Sep 2024) | Google DeepMind | 76.6% | cited-phi35moe | official ↗ |
| #4 | Gemma 2 9B | Google DeepMind | 76.6% | instruct-cited-phi35moe | official ↗ |
| #5 | Qwen2.5 Instruct (72B) | Qwen | 69.9% | instruct-cited-tulu3 | official ↗ |
| #6 | Qwen2.5-7B | Qwen | 69.4% | cited-phi4-mini | official ↗ |
| #7 | Llama 3.1-8B | Meta AI | 69.2% | instruct-cited-phi35moe | official ↗ |
| #8 | Mistral NeMo | Mistral AI | 68.1% | instruct-cited-phi35moe | official ↗ |
| #9 | Granite 3.2 8B | IBM | 66.92% | instruct-from-hf-readme | official ↗ |
| #10 | Llama 3.1-70B | Meta AI | 66.8% | instruct-cited-tulu3 | official ↗ |
| #11 | Hermes 3 70B | Nous Research | 66.5% | cited-tulu3 | official ↗ |
| #12 | Phi-4 Mini | Microsoft | 66.4% | instruct-mc2-10-shot | official ↗ |
| #13 | Granite 3.1 8B | IBM | 65.84% | instruct-cited-granite-3-2 | official ↗ |
| #14 | phi-3.5-mini | Microsoft | 65.2% | cited-phi4-mini | official ↗ |
| #15 | Qwen2.5-3B | Qwen | 64.3% | cited-phi4-mini | official ↗ |
| #16 | Qwen2.5-7B | Qwen | 63.06% | instruct-cited-granite-3-2 | official ↗ |
| #17 | Ministral 8B | Mistral AI | 62.9% | cited-phi4-mini | official ↗ |
| #18 | Gemma 2 9B | Google DeepMind | 61.4% | instruct-cited-tulu3 | official ↗ |
| #19 | Granite 3.2 2B | IBM | 59.8% | instruct-from-hf-readme | official ↗ |
| #20 | Granite 3.1 2B | IBM | 59.79% | instruct-cited-granite-3-2 | official ↗ |
| #21 | Llama 3.2 3B | Meta AI | 59.2% | cited-phi4-mini | official ↗ |
| #22 | Qwen2-57B-A14B | Qwen | 57.7% | base-from-hf-readme | official ↗ |
| #23 | Qwen1.5-32B | Qwen | 57.4% | cited-qwen2-57b | official ↗ |
| #24 | Ministral 8B | Mistral AI | 55.5% | instruct-cited-tulu3 | official ↗ |
| #25 | Llama 3.1-8B | Meta AI | 55.1% | instruct-cited-tulu3 | official ↗ |
| #26 | Tulu 3 8B | Allen Institute for AI | 55% | instruct-6-shot | official ↗ |
| #27 | Qwen2-7B | Qwen | 54.2% | base-from-hf-readme | official ↗ |
| #28 | Yi-1.5-34B | 01.AI | 53.9% | cited-qwen2-57b | official ↗ |
| #29 | Falcon Mamba | Technology Innovation Institute | 53.42% | base-from-hf-readme | official ↗ |
| #30 | Llama 3.1-8B | Meta AI | 52.79% | instruct-cited-granite-3-2 | official ↗ |
| #31 | Mixtral 8x7B | Mistral AI | 51.1% | cited-qwen2-57b | official ↗ |
| #32 | DeepSeek-R1-Distill-Llama-8B | DeepSeek | 47.43% | cited-granite-3-2 | official ↗ |
| #33 | DeepSeek-R1-Distill-Qwen-7B | Qwen | 47.14% | cited-granite-3-2 | official ↗ |
| #34 | DeepSeek-R1-Distill-Qwen-7B | DeepSeek | 47.14% | cited-granite-3-2 | official ↗ |
| #35 | Jamba | AI21 Labs | 46.4% | cited-qwen2-57b | official ↗ |
| #36 | Qwen2-1.5B | Qwen | 45.9% | base-from-hf-readme | official ↗ |
| #37 | Gemma 7B | Google DeepMind | 44.79% | cited-falcon-mamba | official ↗ |
| #38 | Phi-2 | Microsoft | 44.5% | cited-qwen2 | official ↗ |
| #39 | Llama 3.1-8B | Meta AI | 44.29% | base-cited-falcon-mamba | official ↗ |
| #40 | Llama 3-8B | Meta AI | 42.93% | base-cited-falcon-mamba | official ↗ |
| #41 | Mistral 7B | Mistral AI | 42.15% | v0-1-cited-falcon-mamba | official ↗ |
| #42 | Qwen2-0.5B | Qwen | 39.7% | base-from-hf-readme | official ↗ |
| #43 | Gemma 2B | Google DeepMind | 33.1% | cited-qwen2 | official ↗ |
Frequently asked questions about TruthfulQA
What is the TruthfulQA benchmark?
817 questions designed to elicit common human misconceptions and test factual honesty.
How is the TruthfulQA benchmark scored?
TruthfulQA 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 TruthfulQA?
As of the latest reported scores on GenAIList, GPT-4o mini achieves the highest result on TruthfulQA with a score of 78.2%.
Is a higher TruthfulQA score better?
Yes. On TruthfulQA a higher score indicates better performance, so models near the top of the leaderboard are the strongest.