deepseek-v3.2-thinking

DeepSeek-V3.2 is a major large language model release scheduled for December 1, 2025. It comes with both a formal version optimized for everyday applications and a Specialized version enhanced for long-form reasoning. This model achieves a key breakthrough in inference capability, reaching GPT-5-level performance, and for the first time integrates reasoning with tool-call capabilities. Thanks to its open-source nature and rich feature set, it is well-suited for diverse scenarios. Meanwhile, the Specialized version has achieved outstanding results in top international competitions, pushing the performance of open-source models to a new level.

Basic Information

• Released simultaneously with DeepSeek-V3.2-Specialized on December 1, 2025; the former is the official version optimized for everyday applications, while the latter is the long-form reasoning-enhanced edition.
• The official website, app, and API services have all been fully updated to version V3.2. Both models are open-sourced on the HuggingFace and ModelScope platforms.
• The Specialized version is available via a temporary API until December 15, 2025, with a default maximum output length of 128K and supports only reasoning-mode conversations.

Core Features

• Inference capability reaches GPT-5 level; its performance on mainstream inference benchmarks is slightly below Gemini-3.0-Pro, yet its output length is significantly shorter than Kimi-K2-Thinking.
• It’s the first model to integrate reasoning with tool calls, supporting dual-mode tool invocation—both reasoning and non-reasoning modes—and adding support for Claude Code.
• The Specialized version won gold medals in four top international competitions, including IMO 2025 and CMO 2025. Its performance in ICPC and IOI ranked second and tenth, respectively, among human competitors.

Technical Highlights

• Adopts the DSA sparse attention mechanism, which has been validated by the V3.2-Exp experimental version and shows no performance degradation compared to V3.1-Terminus in specific scenarios.
• Constructed over 1,800 environments and more than 85,000 complex instructions for reinforcement learning tasks, enabling large-scale agent training data synthesis.
• The Specialized version incorporates DeepSeek-Math-V2’s theorem-proving capabilities, giving it rigorous mathematical proof and logical verification abilities.

Market Impact

• The model achieves the highest evaluation scores among open-source models, significantly narrowing the gap between open-source and closed-source models in tool-call capabilities.
• The Specialized version delivers groundbreaking performance in top international competitions, driving the extreme enhancement of open-source models’ inference capabilities.

Application Scenarios

• The V3.2 version is suitable for question-answering scenarios and general agent tasks, balancing inference capability and efficiency while reducing computational costs.
• The Specialized version is ideal for high-end research scenarios such as mathematical proofs and logical verification; it is not yet optimized for everyday conversations or writing tasks.
• Supports multi-scenario tool calls; in reasoning mode, multiple rounds of thinking combined with tool calls can produce more accurate responses.

Huggingface open-source link: https://huggingface.co/deepseek-ai/DeepSeek-V3.2

ModelScope (Mogu Community) open-source link: https://modelscope.cn/models/deepseek-ai/DeepSeek-V3.2

Tips:

• If you’d like to use the Claude code format, simply change the model name to deepseek-reasoner in Claude code, or enable reasoning mode by pressing the Tab key in the Claude Code CLI.
• If you’d like to experience the ultimate performance of the Specialized version, please visit deepseek-v3.2-speciale