DeepSeek V4 Pro Preview: benchmark results
DeepSeek's preview release of V4 Pro (April 2026), a 1.6-trillion-parameter, 49-billion-active Mixture-of-Experts model supporting a 1-million-token context, later superseded by the DeepSeek-V4-Pro-0813 official release. Provider: DeepSeek. Released 2026-04-24. Access: Open.
Aggregate Rating 1728 ± 18, rank #148 of 1656 rated models, from 28 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| LLM2014 Logic 2026-07 | 49.97 | Median Score | 74.4 |
| CyberGym | 57.7 | Level 1 successful PoC rate (%) (self-reported) | 46.5 |
| NYT Connections Extended | 67.3 | Score (%) | 46.5 |
| KernelBench Mega | 4.64 | Speedup (x) | 46.4 |
| MineBench | 1469 | Elo Rating | 44.6 |
| Creative Writing (Lechmazur) | -0.54 | Mean Score | 41.4 |
| Lech Mazur Writing | -0.54 | Comparison Score (points) | 41.4 |
| SuperCLUE-Law - Overall | 78.7 | Score | 40 |
| E-Commerce Bench | 190 | Final Assets (¥k, mean of 5 episodes) | 32.5 |
| Multi-turn Debate (Lechmazur) | 1443.6 | Bradley-Terry Rating | 31 |
| SuperCLUE-Terminal - Overall | 39.39 | Score | 29.4 |
| NL2Repo | 38.5 | Score (mixed sources) | 26.1 |
Interactive version: theaggregate.ai/model?slug=deepseek-v4-pro-preview · How It Works · Data refreshed daily, snapshot 2026-10-20.