DeepSeek V4 Pro (0813): benchmark results
DeepSeek's August 13, 2026 official release of V4 Pro, an open-weight MoE reasoning model that supersedes the earlier V4 Pro preview with a DSpark speculative-decoding module for stronger agentic performance. Provider: DeepSeek. Released 2026-08-13. Access: Open.
Aggregate Rating 1819 ± 7, rank #53 of 1656 rated models, from 451 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| DeepSeek MRCR 1M (BenchLM) | 83.5 | MMR (%) | 100 |
| EnterpriseOps-Gym-AA - Drive | 66.67 | Task Success (%) | 100 |
| EnterpriseOps-Gym-AA - Hybrid | 45.75 | Task Success (%) | 100 |
| IndicBankBench | 58.2 | Strict pass^3 (%; cases passed out of all 799 cases (twelve | 100 |
| IndicBankBench - Third-Party Request Refusal | 100 | Strict pass^3 (%; 13 cases passed out of the 13 cases of the | 100 |
| LLM Stats (AutomationBench; Public (DeepSeek minimal harness)) | 31.8 | Score (%) | 100 |
| SuperCLUE General (July 2026) - Hallucination Control | 89.73 | Score | 100 |
| AA Omniscience - Software Engineering (SWE) - R | 64 | Accuracy (%) | 99.7 |
| AA Omniscience - Software Engineering (SWE) - Go | 66 | Accuracy (%) | 99.5 |
| AA Omniscience - Software Engineering (SWE) - PHP | 78 | Accuracy (%) | 99.1 |
| AA Omniscience - Software Engineering (SWE) - TypeScript | 71.11 | Accuracy (%) | 99.1 |
| Dnotitia Korean LLM Leaderboard | 99.73 | Mean Item Score (%) | 99.1 |
Interactive version: theaggregate.ai/model?slug=deepseek-v4-pro-0813 · How It Works · Data refreshed daily, snapshot 2026-10-20.