1.google/Gemma-4-31B-IT-NVFP4
exact_match97.0%±0.5%
GSM8K
MathReasoning
GSM8K (Grade School Math 8K) is a dataset of 8,500 linguistically diverse English-language grade-school math word problems, split into 7,473 training and 1,319 test items, each requiring between two and eight elementary arithmetic steps to solve. The problems were authored by professional human problem-writers, with reference solutions that include natural-language reasoning chains in addition to the final numeric answer. GSM8K became the standard probe for whether a language model can chain together multiple arithmetic steps under natural- language framing, and it remains a near-mandatory entry on foundation-model evaluation cards.
Source paperLatest run: 2026-05-19
Benchmark results
Switch between the canonical ranking, release-date performance view, and score-size tradeoff.
28 models
2.google/gemma-4-31B-it
exact_match96.6%±0.5%
3.google/gemma-4-26B-A4B-it
exact_match95.6%±0.6%
4.Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8
exact_match94.2%±0.6%
5.Qwen/Qwen3-32B
exact_match94.0%±0.7%
6.google/gemma-3-27b-it
exact_match88.5%±0.9%
7.Qwen/Qwen3-14B
exact_match87.6%±0.9%
8.google/gemma-3-12b-it
exact_match86.3%±0.9%
9.nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4
exact_match86.1%±1.0%
10.Qwen/Qwen3-8B
exact_match84.9%±1.0%
11.Qwen/Qwen3.5-122B-A10B-NVFP4
exact_match81.5%±1.1%
12.Qwen/Qwen3-235B-A22B-Thinking-AWQ-2507
exact_match80.7%±1.1%
13.zai-org/GLM-4.5-Air-FP8
exact_match80.0%±1.1%
14.Qwen/Qwen3.5-35B-A3B
exact_match79.5%±1.1%
15.openai/gpt-oss-120b
exact_match74.5%±1.2%
16.Qwen/Qwen3-Next-80B-A3B-Instruct
exact_match73.1%±1.2%
17.Qwen/Qwen3-4B
exact_match71.8%±1.2%
18.MiniMax/MiniMax-M2.1-AWQ
exact_match70.4%±1.3%
19.Qwen/Qwen3.6-27B
exact_match69.1%±1.3%
20.Qwen/Qwen3-4B-AWQ
exact_match62.4%±1.3%
21.openai/gpt-oss-20b
exact_match39.3%±1.3%
22.microsoft/phi-4-mini-instruct
exact_match37.1%±1.3%
23.MiniMax/MiniMax-M2-AWQ
exact_match28.8%±1.2%
24.microsoft/phi-4
exact_match9.7%±0.8%
25.microsoft/phi-4-reasoning-plus
exact_match9.2%±0.8%
26.microsoft/phi-4-mini-reasoning
exact_match4.3%±0.6%
27.zai-org/GLM-4.5V-FP8
exact_match3.5%±0.5%
28.Qwen/Qwen3.6-35B-A3B
exact_match1.1%±0.3%
Caveats
Frontier models now score above 95% on GSM8K, so the benchmark no longer differentiates leading systems and serves mostly as a sanity check that a model can do basic chained arithmetic at all. Beyond saturation, multiple lines of evidence suggest reported scores overstate genuine reasoning ability. Mirzadeh et al. (2024, GSM-Symbolic) showed that simply renaming variables or perturbing the numeric values in GSM8K problems causes large accuracy drops and high variance across frontier models, consistent with template matching rather than reasoning. Zhang et al. (2024, GSM1k) constructed a contamination-free parallel set in the same style and observed accuracy drops of up to ~8pp for some model families, indicating systematic overfitting to the public test split. The scope is also narrow: GSM8K is arithmetic-only and does not probe symbolic, algebraic, or proof-style mathematical reasoning, and it is English-only with no multilingual variant inside the dataset itself (MGSM exists for that purpose). High GSM8K scores should not be read as evidence of mathematical capability beyond chained arithmetic.
How to cite
Citation
FrozeBench. "GSM8K." https://frozebench.com/benchmarks/gsm8k. Retrieved 2026-06-04.
BibTeX
@misc{frozebench_gsm8k,
title = {GSM8K},
howpublished = {\url{https://frozebench.com/benchmarks/gsm8k}},
year = {2026},
note = {FrozeBench. Retrieved 2026-06-04.}
}URL
https://frozebench.com/benchmarks/gsm8k