DeepSeek V3.2: benchmark results
Integrates reasoning into tool-use workflows, with tools available in both thinking and non-thinking modes. Agent training uses synthesized tasks across interactive environments. Provider: DeepSeek. Released 2025-12-01. Access: Open.
Aggregate Rating 1695 ± 2, rank #205 of 1656 rated models, from 2580 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| "Newspaper Eat" Means "Not Tasty": A Taxonomy and Benchmark for Coded Language in Real-World Chinese Online Reviews | 78 | F1 (%) | 100 |
| 6G-Bench - Trust, Security and SLA Awareness | 83.8 | Accuracy (%) on the trust, security and SLA awareness tasks | 100 |
| AJ-Bench (Agent-as-a-Judge) | 77.34 | Overall F1 (%), mean of the seven subdomain F1 scores, with | 100 |
| AJ-Bench (Agent-as-a-Judge) - Deep Search | 82.14 | F1 (%) on Deep search (Mind2Web 2-derived items; 52 tasks, 1 | 100 |
| 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) - PPT | 83.14 | F1 (%) on PowerPoint GUI tasks (OSWorld; 21 tasks, 42 trajec | 100 |
| AJ-Bench (Agent-as-a-Judge) - Postgres | 74.13 | F1 (%) on Postgres data-system tasks (MCPMark; 18 tasks, 100 | 100 |
| AJ-Bench (Agent-as-a-Judge) - Word | 78.64 | F1 (%) on Word GUI tasks (OSWorld; 12 tasks, 24 trajectories | 100 |
| ASTRA-bench - Contextual Question Answering | 95.14 | Task-level end-goal success (0-1, shown times 100) on the 1, | 100 |
| ASTRA-bench - Entity Creation | 91.61 | Task-level end-goal success (0-1, shown times 100) on the 1, | 100 |
| ASTRA-bench - Information Retrieval Recall | 95.16 | Step-level information-retrieval recall (0-1, shown times 10 | 100 |
| ASTRA-bench - Milestone Similarity | 93.4 | Mean milestone similarity (0-1, shown times 100) of the agen | 100 |
Interactive version: theaggregate.ai/model?slug=deepseek-v3-2 · How It Works · Data refreshed daily, snapshot 2026-10-20.