DeepSeek V3.2 Exp (Non-reasoning): benchmark results
DeepSeek V3.2 Exp evaluated with reasoning disabled. Provider: DeepSeek. Released 2025-09-29. Access: Open.
Aggregate Rating 1598 ± 14, rank #721 of 2168 rated models, from 92 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| UGI Leaderboard | 54.67 | UGI Score | 96 |
| UGI - Natural Intelligence | 47.5 | NatInt Score | 87.6 |
| UGI - Writing | 51.04 | Writing Score | 86.9 |
| AA Omniscience - Software Engineering (SWE) - Go | 32 | Accuracy (%) | 85.5 |
| AA MMLU-Pro | 83.64 | Accuracy (%) | 85.3 |
| AA Omniscience - Software Engineering (SWE) - R | 22 | Accuracy (%) | 81.4 |
| AA Omniscience - Software Engineering (SWE) - Kotlin | 28 | Accuracy (%) | 80.5 |
| AA Omniscience - Software Engineering (SWE) - Python | 29.5 | Accuracy (%) | 80.2 |
| AA Omniscience - Software Engineering (SWE) - PHP | 34 | Accuracy (%) | 79.4 |
| AgentGym2 - Deep Search | 24.41 | Task success (%; Avg@3 on the 198 deep search tasks; mean ov | 78.6 |
| AgentNoiseBench (Search, Tool Noise) | 36 | Accuracy on 500 sampled 2WikiMultiHopQA and HotpotQA questio | 78.3 |
| AA Omniscience - Software Engineering (SWE) - Java | 21 | Accuracy (%) | 77.3 |
Interactive version: theaggregate.ai/model?slug=deepseek-v3-2-exp-non-reasoning · How It Works · Data refreshed daily, snapshot 2026-10-20.