DeepSeek-V3.2 is a reasoning-first large language model launched by DeepSeek AI, specifically built for agent scenarios. While maintaining high performance, it optimizes computational efficiency through an innovative sparse attention mechanism (DeepSeek Sparse Attention, DSA) and demonstrates excellent performance in complex tasks and agent capabilities. DeepSeek-V3.2 aims to be a powerful foundation for developers to build next-generation AI applications.
- Leading Reasoning Capabilities: DeepSeek-V3.2 performs exceptionally well in multiple benchmarks, especially in Agentic Capabilities. Its Thinking mode is comparable to, or even surpasses, Gemini-3.0-Pro in coding tasks like Codeforces and SWE Verified.
- Sparse Attention Mechanism (DSA): Introduces DeepSeek Sparse Attention (DSA) to optimize the efficiency and speed of long-context reasoning, making it more efficient in processing long texts.
- Agent Scenario Optimization: Through a large-scale agent training data synthesis method covering 1800+ environments and 85k+ complex instructions, it integrates thinking directly into tool-use, supporting tool calls in both thinking and non-thinking modes.
- Context Window: Supports a context length of 164K tokens, capable of processing and understanding large amounts of information.
- Maximum Output: The maximum output can reach 164K tokens.
- Agent Development: Building AI agents capable of complex planning, multi-step execution, and tool calling.
- Code Generation and Assistance: Providing powerful support in software development, code review, and automated programming tasks.
- Long Text Understanding and Summarization: Processing and analyzing ultra-long text content such as legal documents, research reports, and technical manuals.
- Complex Problem Solving: Delivering high-quality solutions in scenarios requiring deep reasoning and logical analysis.
| Capability | Detailed Description |
|---|
| Reasoning Ability | Extremely Strong. Excels in complex logical reasoning, agent task planning, and execution, especially in Thinking mode. |
| Creative Ability | Strong. Proficient in code generation, text creation, and agent task design. |
| Multimodal Ability | Primarily focused on text, but its agent capabilities can be combined with multimodal tools to achieve multimodal tasks. |
| Response Speed | Optimized for long-context reasoning, offering high efficiency. |
| Context Window | 164K Tokens |
| Max Output | 164K Tokens |
| Model | Input (Credits/Token) | Cache Write (Credits/Token) | Cache Read (Credits/Token) | Output (Credits/Token) |
|---|
| DeepSeek-V3.2 | 0.29 | 0.29 | 0.145 | 0.44 |
