DeepSeek Releases V3.2 / V3.2-Speciale

DeepSeek officially launched V3.2 and its enhanced version V3.2-Speciale on December 1, 2025. The release features efficient reasoning and agent capabilities, introduces sparse attention (DSA) to optimize long contexts and reduce computational complexity, and implements the ability to use tools during the 'thinking process' for the first time, supporting both thinking and non-thinking modes for tool invocation. The model enhances generalization in large-scale agent task synthesis and integrates reasoning into tool invocation scenarios. The team trained the new model using over 1,800 complex task environments and 85,000 'hard to answer but easy to verify' instructions. Compared to V3.1, the output is shorter, faster, and consumes less computational power. The long-thinking enhanced version, DeepSeek-V3.2-Speciale, incorporates theorem proving capabilities from DeepSeek-Math-V2 and has achieved notable results in multiple international competitions (including IMO, CMO, ICPC, and IOI). The official assessment states that V3.2 is close to GPT-5, second only to Gemini 3 Pro, while Speciale performs exceptionally in competitions. The official website, app, and API have all launched V3.2, with new Chat Template and developer role support for 'thinking with tools.' The API now includes Claude Code support, and all platforms have been upgraded to the official version. DeepSeek-V3.2-Speciale is available for research purposes only, supporting 'thinking mode' but not tool invocation. Speciale can output up to 128K tokens, with free testing available until December 15, 2025. The release includes model download/access links (multiple public links) and several introductory posts showcasing model capabilities and evaluation results.
Key insights
DeepSeek Releases V3.2 and Introduces Sparse Attention and Long Context Optimization
DeepSeek officially launched V3.2 on December 1, 2025, with major improvements including the introduction of a sparse attention mechanism (DSA) to reduce computational complexity while maintaining performance, and optimizations for long context scenarios. Official evaluations show that V3.2 produces shorter outputs, responds faster, and consumes less computational power compared to its predecessor. The official website, app, and API have all been updated to the official version. Related posts list model download and access links, showcasing images and evaluation summaries.
V3.2 Achieves Agent Capability to Use Tools During the 'Thinking Process'
DeepSeek-V3.2 has implemented the ability for the model to autonomously decide and invoke external tools within the 'thinking chain': during reasoning, it can choose to search the internet, call a code interpreter, or validate through drawing tools, thereby solving multi-step problems in complex interactive environments. The team enhanced generalization through large-scale agent task synthesis training, achieving top performance levels for open-source models even without specialized training on the testing tools.
DeepSeek-V3.2-Speciale Integrates Mathematical Proof Module and Achieves High Rankings in Competitions
DeepSeek simultaneously released the enhanced version V3.2-Speciale, integrating the previous DeepSeek-Math-V2 for long thinking and theorem proving. Speciale has been listed in multiple reports for its achievements in the 2025 IMO, CMO, ICPC World Finals, and IOI, including records such as IMO gold medal, CMO gold medal, second place for human competitors in ICPC, and various rankings in IOI. Speciale is designated as a research-use version, supporting ultra-long thinking chains but not tool invocation.
Training Methodology and Data: 1,800+ Complex Environments and 85,000 Verification Instructions
The DeepSeek team designed a new training method for V3.2, simulating over 1,800 complex task environments and constructing 85,000 'hard to answer but easy to verify' instructions. The goal was to teach the model self-reasoning rather than merely reciting existing knowledge. This training method was used to evaluate the model's ability to generalize to unseen tasks, demonstrating strong generalization performance in official evaluations.
Product and Release Details: API Support, Chat Template, and Open Policy
The release information indicates that the official team has updated the Chat Template to support the thinking with tools capability and introduced a developer role for search agents. The API has added Claude Code support. DeepSeek-V3.2-Speciale is for research purposes only, providing dialogue in thinking mode without tool invocation; Speciale's maximum output can reach 128K tokens, with a free testing period until December 15, 2025. Multiple posts include model access links and showcase images.
Community and Public Opinion: Competition Results Spark Discussion on Practical Application Value
Several posts report Speciale's outstanding achievements in international competitions, while users have raised discussions about the value of competition gold medal models in real-world scenarios, emphasizing the need to test economic value use cases in real life. Comments mention the strongest use cases for different models (writing, agentic search, knowledge comics, etc.) and call for other teams to disclose their competition performance models.