DeepSeek V3.2 (Thinking): benchmark results
DeepSeek V3.2 evaluated with thinking enabled. Provider: DeepSeek. Released 2025-12-01. Access: Open.
Aggregate Rating 1673 ± 5, rank #444 of 2168 rated models, from 688 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| AJ-Bench (Agent-as-a-Judge) - Excel | 86.49 | F1 (%) on Excel GUI tasks (OSWorld; 19 tasks, 38 trajectorie | 100 |
| AJ-Bench (Agent-as-a-Judge) - Postgres | 74.13 | F1 (%) on Postgres data-system tasks (MCPMark; 18 tasks, 100 | 100 |
| AgentWebBench (Centralized Retrieval) - Deep Research | 66.54 | Key Point Recall (%) of long reports on 331 DeepResearchGym | 100 |
| AlpsBench - Constraint Following | 94.19 | Pass rate (0-1, times 100) of responses to user queries that | 100 |
| AlpsBench - Emotional Intelligence (Chinese) | 4 | Emotional-intelligence score (1 to 5) of responses to Chines | 100 |
| AlpsBench - Extraction | 67.42 | Semantic-matching F1 (0-1, times 100) of the structured memo | 100 |
| AlpsBench - Retrieval (100 Distractors) | 97.28 | Recall (0-1, times 100): share of queries for which the mode | 100 |
| AlpsBench - Retrieval (1000 Distractors) | 94.95 | Recall (0-1, times 100): share of queries for which the mode | 100 |
| G-IdiomAlign - Gloss-Contrastive Generation (No Gloss) | 24.66 | Acc@0.80 (%; the model writes one target-language idiom for | 100 |
| G-IdiomAlign - Gloss-Contrastive Generation (With Gloss) | 29.63 | Acc@0.80 (%; the model writes one target-language idiom for | 100 |
| G-IdiomAlign - Multiple-Choice Equivalence - Chinese Targets | 69.37 | Accuracy (%; source idiom to its meaning-equivalent target-l | 100 |
| G-IdiomAlign - Multiple-Choice Equivalence - English Targets | 66.7 | Accuracy (%; source idiom to its meaning-equivalent target-l | 100 |
Interactive version: theaggregate.ai/model?slug=deepseek-v3-2-thinking · How It Works · Data refreshed daily, snapshot 2026-10-20.