Scores updated · 34 tests both models report · How we compare
Every test, side by side
All 34 tests both models report. The winning score is in its model's colour; marks a score checked independently.
Other results34 tests
- FACTS Parametric (EM)DeepSeek-V4-Pro-Base by 28.733.962.6+28.7
- Simple-QA verified (EM)DeepSeek-V4-Pro-Base by 25.130.155.2+25.1
- MultiLoKo (EM)DeepSeek-V4-Pro-Base by 8.942.251.1+8.9
- MultiLoKo (LLM-Judge)DeepSeek-V4-Pro-Base by 8.342.650.9+8.3
- SuperGPQA (EM)DeepSeek-V4-Pro-Base by 7.446.553.9+7.4
- HumanEvalDeepSeek-V4-Pro-Base by 7.369.576.8+7.3
- MATHDeepSeek-V4-Pro-Base by 7.157.464.5+7.1
- MATH (EM)DeepSeek-V4-Pro-Base by 7.157.464.5+7.1
- LongBench-V2 (EM)DeepSeek-V4-Pro-Base by 6.844.751.5+6.8
- MMLU-ProDeepSeek-V4-Pro-Base by 5.268.373.5+5.2
- BIG-Bench Extra HardDeepSeek-V4-Pro-Base by 4.425.429.8+4.4
- CLUEWSCDeepSeek-V4-Pro-Base by 382.285.2+3
- TriviaQADeepSeek-V4-Pro-Base by 2.882.885.6+2.8
- TriviaQA (EM)DeepSeek-V4-Pro-Base by 2.882.885.6+2.8
- CMath (EM)DeepSeek-V4-Flash-Base by 2.793.690.9+2.7
- BigCodeBenchDeepSeek-V4-Pro-Base by 2.456.859.2+2.4
- HellaSwagDeepSeek-V4-Pro-Base by 2.385.788+2.3
- HellaSwag (EM)DeepSeek-V4-Pro-Base by 2.385.788+2.3
- WinoGrandeDeepSeek-V4-Pro-Base by 279.581.5+2
- WinoGrande (EM)DeepSeek-V4-Pro-Base by 279.581.5+2
- GSM8KDeepSeek-V4-Pro-Base by 1.890.892.6+1.8
- GSM8K (EM)DeepSeek-V4-Pro-Base by 1.890.892.6+1.8
- MMMLU (EM)DeepSeek-V4-Pro-Base by 1.588.890.3+1.5
- MMLUDeepSeek-V4-Pro-Base by 1.488.790.1+1.4
- MMLU-ReduxDeepSeek-V4-Pro-Base by 1.489.490.8+1.4
- MGSM (EM)DeepSeek-V4-Flash-Base by 1.385.784.4+1.3
- C-EvalDeepSeek-V4-Pro-Base by 192.193.1+1
- BBHtie86.987.5tie
- BBH (EM)tie86.987.5tie
- AGIEval (EM)tie83.984.4tie
- CMMLUtie90.490.8tie
- CMMLU (EM)tie90.490.8tie
- DROPtie88.688.7tie
- DROP (F1)tie88.688.7tie
Questions people ask
How do you compare the two?
We use the 34 benchmark tests both models have published scores on. None of them is in the eight capability areas we count, so this page lists them without a verdict. Each score is the one shown on the model's own page.