DeepSeek V3.2 Speciale
DeepSeek V3.2 Speciale, developed by DeepSeek-AI, is a reasoning-focused foundation model with high computational capability. By relaxing output limits, it enhances complex reasoning abilities and achieves gold medal-level performance in top-level competitive programming such as the International Mathematical Olympiad and Informatics Olympiad.
Parameters
Undisclosed
Context Window
License
https://github.com/deepseek-ai/deepseek-LLM/blob/main/LICENSE-MODEL
Release Date
2025-12-01
API Pricing
API pricing for this model is not yet available
Strengths
- ・Extremely high complex reasoning ability
- ・Strong in competitive programming
- ・Performance surpassing GPT-5
Weaknesses
- ・Not compatible with tool calling
- ・Requires high computational resources
- ・Lack of general-purpose feature expansion
Use Cases
- ・Solving difficult mathematical problems
- ・Advanced code generation
- ・Competitive programming preparation
Deep Analysis
Parameters
685B MoE
Arena Elo
1361
#7 overall
AIME 2025
96.0%
HMMT Feb 2025
99.2%
Codeforces
2701
IMO 2025
Gold Medal
Release Date
December 1, 2025
Strengths
- ・IMO 2025 and IOI 2025 gold medals
- ・Surpasses GPT-5 on multiple benchmarks
- ・Fully open-source MIT
- ・#7 on Chatbot Arena
Weaknesses
- ・High token consumption (23-77k per query)
- ・Retired from DeepSeek API
- ・Slower due to extended reasoning
- ・Weaker on general conversational tasks
Competitor Comparison
| Model | Arena | GPQA | Price |
|---|---|---|---|
| GPT-5 High | ~1380 | 85.7% | $2.50/$15 |
| Gemini 3.0 Pro | ~1389 | 91.9% | |
| DeepSeek-V3.2 | 1334 | 82.4% | $0.28/$0.42 |
DeepSeek V3.2 Speciale is the most powerful reasoning model. IMO/IOI gold medals, 96% AIME, #7 Arena Elo 1361.
Sources
Analysis generated: 2026-05-24