Across 14 shared benchmarks, DeepSeek-V3 scores higher on 13 and ERNIE 4.5 on 1. The widest gap is MMLU-Pro, where DeepSeek-V3 scores 81.2 against 16.
| Benchmark | DeepSeek-V3 | ERNIE 4.5 |
|---|---|---|
| arc_challenge | 95.3 | 40.6 |
| bbh | 87.5 | 30.4 |
| C-Eval | 90.1 | 40.7 |
| CLUEWSC | 90.9 | 48.6 |
| cmmlu | 88.8 | 39.8 |
| DROP | 91.6 | 28.6 |
| GPQA Diamond | 68.4 | 74 |
| GSM8K | 96.7 | 25.2 |
| hellaswag | 88.9 | 33 |
| mmlu_redux | 90.5 | 43.2 |
| MMLU-Pro | 81.2 | 16 |
| PIQA | 84.7 | 55.2 |
| simpleqa | 27.7 | 1.8 |
| winogrande | 84.9 | 51.3 |
Best tracked score per model per benchmark (default configuration; source-attributed). ↓ marks lower-is-better metrics. Open either model for its full surface, provenance and pricing.