DeepSeek-V3.2 (DSA, scaled RL, agentic synthesis)
Pairs sparse attention with a scaled RL budget and a synthetic agentic-task pipeline; V3.2-Speciale claims IMO and IOI 2025 gold-level results.
It makes three claims. DSA keeps long-context quality at lower cost; post-training RL compute scaled to GPT-5-class performance; a pipeline synthesises tool-use environments. Speciale (high-compute variant) is reported to match Gemini 3.0 Pro reasoning. Company-reported.
- Date
- Tuesday, 2 December 2025
- Lab
- DeepSeek
- Kind
- paper
- Access
- open weights
arXiv v1 2025-12-02; 264+ authors. Competition-gold and 'GPT-5 parity' claims are from the abstract (company-reported) and were not independently verified here.
Sources
This record was checked against its sources on 6 October 2026. How we check