DeepSeek-V3.2
Overviewβ
DeepSeek-V3.2 is a reasoning-first large language model launched by DeepSeek AI and designed for agent scenarios. While maintaining strong task performance, it improves long-context efficiency through the DeepSeek Sparse Attention (DSA) mechanism and supports tool-oriented workflows in both thinking and non-thinking modes.
π Key Featuresβ
- Strong Reasoning for Agent Workflows: Performs well on complex reasoning, coding, and tool-oriented agent tasks, with a dedicated Thinking mode for deeper task planning and execution.
- DeepSeek Sparse Attention (DSA): Introduces DeepSeek Sparse Attention (DSA) to improve the efficiency and speed of long-context reasoning.
- Optimized for Agent Scenarios:
- Environment Coverage: Trained with data spanning 1,800+ environments.
- Complex Instruction Following: Refined with 85k+ complex instructions.
- Native Tool Integration: Supports tool calling in both Thinking and non-Thinking modes.
- Long Context and Large Output: Supports a 164K token context window and up to 164K tokens of output.
π‘ Best Use Casesβ
- Agent Development: Building AI agents capable of complex planning, multi-step execution, and tool calling.
- Code Generation and Assistance: Supporting software development, code review, and automated programming tasks.
- Long Text Understanding and Summarization: Processing large documents such as legal files, research reports, and technical manuals.
- Complex Problem Solving: Handling tasks that require deeper reasoning and structured logical analysis.
π Capabilities and Limitationsβ
| Capability | Detailed Description |
|---|
| Reasoning Ability | 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 can be combined with multimodal tools through agent workflows. |
| Response Speed | Optimized for long-context reasoning, offering high efficiency. |
| Context Window | 164K Tokens |
| Max Output | 164K Tokens |
Credits Usageβ
| Model | Input (Credits/Token) | Cache Write (Credits/Token) | Cache Read (Credits/Token) | Output (Credits/Token) | Web Search (Credits/Use) | Billing Notes |
|---|
| DeepSeek V3.2 | 0.29 | 0.29 | 0.145 | 0.44 | - | - |