Skip to content
other benchmark · included in ECI

PIQA

PIQA benchmark (score column: Score).
Results
34
Random baseline
50.0%
Score ceiling
100%
Released
26 Nov 2019

Score by model release date

Best score by organisation

#ModelScoreRelativeEvidence
1Phi-3.5-MoE Open source
Microsoft
88.6%Source ↗
2Llama 3.1-405B Open source
Meta AI
85.9%Source ↗
3Falcon-180B Open source
Technology Innovation Institute
84.9%Source ↗
4DeepSeek V3 Open source
DeepSeek
84.7%Source ↗
5DeepSeek-V2 (MoE-236B, May 2024) Open source
DeepSeek
83.9%Source ↗
6Gemma 2 9B Open source
Google DeepMind
83.7%Source ↗
7Mixtral 8x7B Open source
Mistral AI
83.6%Source ↗
8Mistral Nemo Open source
Mistral
83.5%Source ↗
9Stable Beluga 2 Open source
Stability AI
83.3%Source ↗
10Mistral 7B v0.1 Open source
Mistral AI
83.0%Source ↗
11Falcon-40B Open source
Technology Innovation Institute
83.0%Source ↗
12Llama 2-70B Open source
Meta AI
82.8%Source ↗
13LLaMA-65B Open source
Meta AI
82.8%Source ↗
14Qwen2.5-72B Open source
Alibaba
82.6%Source ↗
15LLaMA-33B Open source
Meta AI
82.3%Source ↗
16Mistral 7B v0.2 Open source
Mistral AI
82.2%Source ↗
17MPT-30B Open source
MosaicML
81.9%Source ↗
18Llama 3.1-8B Open source
Meta AI
81.2%Source ↗
19Gemma 7B Open source
Google DeepMind
81.2%Source ↗
20Phi-3.5-mini Open source
Microsoft
81.0%Source ↗
21Llama 2-13B Open source
Meta AI
80.8%Source ↗
22MPT-7B Open source
MosaicML
80.6%Source ↗
23Falcon-7B Open source
Technology Innovation Institute
80.3%Source ↗
24LLaMA-13B Open source
Meta AI
80.1%Source ↗
25Qwen-14B Open source
Alibaba
79.9%Source ↗
26LLaMA-7B Open source
Meta AI
79.8%Source ↗
27Llama 2-7B Open source
Meta AI
78.8%Source ↗
28Baichuan2-13B Open source
Baichuan
78.1%Source ↗
29Qwen-7B Open source
Alibaba
77.9%Source ↗
30Gemma 2B Open source
Google DeepMind
77.3%Source ↗
31Baichuan1-7B Open source
Baichuan
76.2%Source ↗
32XGen-7B Open source
Salesforce
75.5%Source ↗
33Dolly 2.0-12b Open source
Databricks
75.4%Source ↗
34Cerebras-GPT-13B Open source
Cerebras Systems
73.5%Source ↗

Caveats: settings (reasoning effort, agent scaffold) differ between rows and materially affect scores; the best reported setting per model is shown. Source: Epoch AI, CC BY 4.0.