DeepSeek V3 (0324): benchmark results
March 24, 2025 snapshot of DeepSeek's open V3 MoE chat model. Provider: DeepSeek. Released 2025-03-24. Access: Open.
Aggregate Rating 1636 ± 5, rank #340 of 1656 rated models, from 792 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| AGC-Bench - pun_eval | 1.52 | Dataset z-score | 100 |
| BenCzechMark | 83.12 | Average Duel Win Score (%) | 100 |
| BenCzechMark - Factual Knowledge | 88.5 | Duel Win Score (%) | 100 |
| CommunityBench - Preference Identification | 40.34 | Accuracy (%) | 100 |
| ControBench - Abortion | 62.94 | Macro-F1 (%; 3-class user stance, one-shot prompting on a st | 100 |
| FinRiskAtlas - Applied Review - Decision-View Generation | 71.31 | Rubric Score (%; DeepSeek-V4-Flash judge) | 100 |
| GitTaskBench (Aider) | 16.67 | Task pass rate (%) | 100 |
| GitTaskBench (Aider) - Execution Completion | 20.37 | Execution completion rate (%) | 100 |
| OpenCompass Code - Execution - Chinese (CompassBench 2504) | 30.4 | Score (%) | 100 |
| PCoA - Summary | 0.66 | Claim F1 (0-1; harmonic mean of claim recall against referen | 100 |
| SWE-bench Live Lite (Agentless) | 13.33 | Resolved rate (%) | 100 |
| Open PL LLM - Belebele (generative, 5-shot) | 93.67 | Accuracy (%) | 99.7 |
Interactive version: theaggregate.ai/model?slug=deepseek-v3-0324 · How It Works · Data refreshed daily, snapshot 2026-10-20.