DeepSeek today released two official AI models, DeepSeek-V3.2 and DeepSeek-V3.2-Speciale. The former targets everyday applications, while the latter has secured gold medals in multiple international competitions, marking a significant step in narrowing the performance gap between open-source and proprietary commercial models.
According to the company, DeepSeek-V3.2 achieves GPT-5-level performance in public reasoning tests, slightly trailing only Gemini-3.0-Pro. Compared to similar products like Kimi-K2-Thinking, the new model significantly reduces output length, lowering computational costs and user wait times. Official web, app, and API services have been fully updated to V3.2.
The enhanced V3.2-Speciale version won gold in four top-tier international competitions, including IMO 2025 and CMO 2025, with ICPC and IOI performances ranking second and tenth among human participants, respectively. This version integrates DeepSeek-Math-V2's theorem-proving capabilities, pushing open-source model reasoning to its limits.
Both models are now open-sourced on HuggingFace and ModelScope. V3.2-Speciale is temporarily available via API until December 15 for community evaluation and research.
**Performance Benchmarking Against Top Proprietary Models** DeepSeek-V3.2 balances reasoning and output length for daily use, excelling in Q&A and general agent tasks. In mainstream reasoning benchmarks, it performs close to Gemini-3.0-Pro.
V3.2-Speciale, optimized for extended reasoning, demonstrates rigorous mathematical proof and logic verification, achieving groundbreaking results in IMO, CMO, ICPC, and IOI—a first for open-source models. However, the company notes its higher token consumption and costs for complex tasks, limiting it to research use without tool-calling support.
**Pioneering Thought-Integrated Tool Usage** DeepSeek-V3.2 is the company’s first model to integrate reasoning with tool usage, supporting both thought and non-thought modes. A novel large-scale agent training data synthesis method was developed, creating 1,800+ environments and 85,000+ complex instructions to enhance generalization.
Technical reports indicate V3.2 leads open-source models in agent evaluations, bridging the tool-calling gap with proprietary models. The model shows strong real-world applicability without test-set-specific training.
In thought mode, multi-step reasoning and tool calls yield more accurate responses. Claude Code support has been added, though compatibility with non-standard tools like Cline and RooCode remains limited.
**DSA Sparse Attention Mechanism Validated** User feedback on the experimental DeepSeek-V3.2-Exp, released two months ago, confirmed the effectiveness of the DSA sparse attention mechanism, showing no significant performance drop versus V3.1-Terminus.
Official platforms now host the finalized V3.2, while V3.2-Speciale’s temporary API (max 128K output) runs until December 15. Both models are open-sourced, with technical reports published. The company credits user engagement for driving innovation.
For details, visit: - [DeepSeek-V3.2](https://huggingface.co/deepseek-ai/DeepSeek-V3.2) - [DeepSeek-V3.2-Speciale](https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Speciale) - [Technical Report](https://modelscope.cn/models/deepseek-ai/DeepSeek-V3.2/resolve/master/assets/paper.pdf)
Comments