1.google/Gemma-4-31B-IT-NVFP4
acc50.7%±2.4%
GPQA Main
KnowledgeScienceReasoning
GPQA Main consists of 448 multiple-choice questions in biology, physics, and chemistry, authored by PhD-level domain experts. It is the base set of the GPQA benchmark, filtered to questions where expert validators demonstrated high accuracy while skilled non-experts with web access failed.
Source paperLatest run: 2026-05-25
Benchmark results
Switch between the canonical ranking, release-date performance view, and score-size tradeoff.
28 models
2.google/gemma-4-31B-it
acc50.4%±2.4%
3.Qwen/Qwen3-Next-80B-A3B-Instruct
acc42.4%±2.3%
4.google/gemma-4-26B-A4B-it
acc41.7%±2.3%
5.google/gemma-3-27b-it
acc38.4%±2.3%
6.google/gemma-3-12b-it
acc35.7%±2.3%
7.microsoft/phi-4-mini-instruct
acc33.7%±2.2%
8.microsoft/phi-4
acc33.5%±2.2%
9.Qwen/Qwen3-4B-AWQ
acc33.0%±2.2%
10.Qwen/Qwen3.6-35B-A3B
acc29.9%±2.2%
11.Qwen/Qwen3-235B-A22B-Thinking-AWQ-2507
acc29.9%±2.2%
12.nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4
acc29.5%±2.2%
13.Qwen/Qwen3-32B
acc28.8%±2.1%
14.Qwen/Qwen3.6-27B
acc28.8%±2.1%
15.Qwen/Qwen3.5-35B-A3B
acc28.3%±2.1%
16.Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8
acc28.3%±2.1%
17.openai/gpt-oss-20b
acc26.8%±2.1%
18.openai/gpt-oss-120b
acc26.1%±2.1%
19.Qwen/Qwen3-14B
acc26.1%±2.1%
20.Qwen/Qwen3.5-122B-A10B-NVFP4
acc26.1%±2.1%
21.microsoft/phi-4-reasoning-plus
acc26.1%±2.1%
22.zai-org/GLM-4.5-Air-FP8
acc25.7%±2.1%
23.Qwen/Qwen3-8B
acc25.4%±2.1%
24.microsoft/phi-4-mini-reasoning
acc25.4%±2.1%
25.Qwen/Qwen3-4B
acc25.2%±2.1%
26.MiniMax/MiniMax-M2.1-AWQ
acc24.3%±2.0%
27.zai-org/GLM-4.5V-FP8
acc24.3%±2.0%
28.MiniMax/MiniMax-M2-AWQ
acc21.4%±1.9%
Caveats
The total set is small (N=448), so confidence intervals on per-model accuracy are wide. The questions were authored by a relatively small pool of domain experts, which introduces stylistic and difficulty biases. Contamination is a concern for models trained after late 2023.
How to cite
Citation
FrozeBench. "GPQA Main." https://frozebench.com/benchmarks/gpqa-main. Retrieved 2026-06-04.
BibTeX
@misc{frozebench_gpqa_main,
title = {GPQA Main},
howpublished = {\url{https://frozebench.com/benchmarks/gpqa-main}},
year = {2026},
note = {FrozeBench. Retrieved 2026-06-04.}
}URL
https://frozebench.com/benchmarks/gpqa-main