DeepSeek-V3.2 and V3.2-Speciale
Production DSA model at claimed GPT-5 level; Speciale variant claims gold at IMO, CMO, ICPC World Finals and IOI 2025. First thinking-in-tool-use release.
DSA plus scaled RL post-training and a synthetic agent-task pipeline (1,800+ environments, 85K+ complex instructions) that lets the model reason while calling tools. Speciale is a high-compute reasoning-only variant on a temporary API endpoint until 2025-12-15.
- Date
- Monday, 1 December 2025
- Lab
- DeepSeek
- Kind
- open-weights
- Access
- open weights
- Price
- Speciale priced same as V3.2
Figures
| Measure | Value | Measured by |
|---|---|---|
| Agent training coverage | 1,800+ environments, 85K+ instructions Speciale is gold-level at IMO, CMO, ICPC WF, IOI 2025 (company claim) and has no tool calling | company |
Parity claims versus GPT-5 and Gemini 3.0 Pro are DeepSeek's own; the paper itself flags independent verification as needed. MIT license; 685B parameters.
Sources
- api-docs.deepseek.com/news/news251201
- arxiv.org/abs/2512.02556
- huggingface.co/deepseek-ai/DeepSeek-V3.2
This record was checked against its sources on 6 October 2026. How we check