DeepSeek V3.2 Speciale: benchmark results
Reasoning-focused V3.2 variant that spends more tokens on difficult problems. Its research API release omitted tool calling, unlike standard V3.2. Provider: DeepSeek. Released 2025-12-01. Access: Open.
Aggregate Rating 1739 ± 11, rank #128 of 1656 rated models, from 195 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| LLM Stats (Codeforces Rating - DeepSeek V3.2 report) | 2701 | Codeforces Rating | 100 |
| MathNet-RAG - Expert Retrieval | 97.3 | on MathNet-RAG, 35 Olympiad problems each paired by experts | 100 |
| OpenCompass Research - AIME 2025 | 96 | Score (%) | 100 |
| SuperCLUE General (November 2025) - Math Reasoning | 75.24 | Score | 100 |
| UGI Leaderboard | 65.37 | UGI Score | 99.5 |
| AA LiveCodeBench | 89.63 | Pass@1 (%) | 99.4 |
| AA AIME 2025 | 96.67 | Accuracy (%) | 98.3 |
| BenchTable - Utility | 82.6 | Weighted Score (%) | 98.1 |
| JudgeBench Reasoning | 96.94 | Accuracy (%) | 98 |
| BenchTable - STEM | 87.1 | Weighted Score (%) | 97.8 |
| MathArena - CMIMC 2025 | 95.62 | Accuracy (%) | 97.1 |
| OpenCompass Research - HLE | 28.6 | Score (%) | 96.8 |
Interactive version: theaggregate.ai/model?slug=deepseek-v3-2-speciale · How It Works · Data refreshed daily, snapshot 2026-10-20.