DeepSeek releases V3.2 and a temporary high-compute V3.2-Speciale

DeepSeek/DeepSeek-V series

DeepSeek released V3.2 as its main model and V3.2-Speciale, a high-reasoning variant available through API until 15 December 2025. Both were open-sourced; V3.2 was trained with synthesised data across 1,800+ agent environments and reached gold-medal level at the 2025 IMO and IOI.

• V3.2-Speciale achieves competitive performance on international competition problems; temporary availability for research evaluation.
• Trained with agent-based synthetic data integrating thinking directly into tool-use workflows.
• Both models share V3.2-Exp pricing and usage patterns; open-source versions available on Hugging Face.