DeepSeek-V3.2 - The official release of DeepSeek's open-source AI model.

What is DeepSeek-V3.2?

DeepSeek-V3.2-Exp is an experimental artificial intelligence model launched by DeepSeek-AI. By introducing the DeepSeek Sparse Attention (DSA) mechanism, it significantly improves the efficiency of long text processing. The model is continuously trained based on DeepSeek-V3.1-Terminus, and only introduces DSA in its architecture to implement a fine-grained sparse attention mechanism. It uses a lightning indexer to efficiently select key information, greatly improving efficiency in long text training and inference.

In terms of performance, DeepSeek-V3.2-Exp is roughly on par with DeepSeek-V3.1-Terminus on multiple public benchmark datasets, demonstrating its capabilities across various domains. The model is open-source on the Hugging Face and ModelScope platforms, facilitating exploration and application by researchers and developers. The significantly reduced API price of DeepSeek-V3.2-Exp lowers the cost for developers, further promoting its widespread deployment in practical applications.

• DeepSeek-V3.2 (Standard Edition)Suitable for everyday use, balancing inference speed and output quality.
• DeepSeek-V3.2-Speciale (Enhanced Version)Optimized for complex reasoning and long thinking tasks, it offers higher performance but at a higher computational cost.

Main functions of DeepSeek-V3.2

• Architectural InnovationDeepSeek-V3.2-Exp introduces the DeepSeek Sparse Attention (DSA) mechanism on the basis of DeepSeek-V3.1-Terminus. Through the lightning indexer and fine-grained label selection mechanism, it achieves significant efficiency improvement, especially in long text scenarios.
• Performance optimizationThe model performs comparably to DeepSeek-V3.1-Terminus on multiple public evaluation sets, significantly reducing inference costs in long text processing. O(L2) Optimized to O(Lk)This significantly improves the efficiency of reasoning in long texts.
• Open source sharingDeepSeek-V3.2-Exp is open source on the Hugging Face and ModelScope platforms, providing detailed implementation details and model weights to facilitate research and application by researchers and developers.
• Cost reductionThe significant drop in API prices has reduced the cost of use for developers, enabling more developers to access and use the model at a lower cost, and promoting its widespread deployment in practical applications.
• Application ExpansionThe official app, web version, and mini-program have all been updated to DeepSeek-V3.2-Exp, supporting multiple application platforms and providing users with a more efficient and economical AI service experience.

Technical Principles of DeepSeek-V3.2

• Reasoning ability greatly improved​
  • Approaching GPT-5 levelIt performs well in multiple public inference benchmarks (such as GSM8K, MATH, HumanEval, etc.), approaching GPT-5, and surpassing Gemini-3.0-Pro in some tasks.
  • Enhanced mathematical abilityIt integrates the theorem proving capabilities of DeepSeek-Math-V2, achieving the level of human gold medalists in top competitions such as the IMO (International Mathematical Olympiad) and ICPC (International Collegiate Programming Contest).
  • More efficient long text processingCompared to its predecessor (V3.1-Terminus), this model reduces redundant content and improves response speed while maintaining high-quality output.
• Significant upgrade to Agent (Intelligent Agent) capabilities​
  • The first model to support the integration of "thinking + tool usage"It can dynamically invoke external tools (such as code execution, search engines, APIs, etc.) during the reasoning process. It supports multi-round thinking modes and performs more stably when solving complex tasks.
  • Large-scale agent training dataTrained on 1800+ different environments and 85,000+ complex instructions, it significantly improves task planning and execution capabilities. It achieves top performance in open-source agent evaluations, narrowing the gap with closed-source models (such as GPT-5 and Claude 4).
• Smarter tool calls​
  • Support Claude Code: Optimizes code execution capabilities and can be combined with tools such as Python and SQL for complex calculations.
  • Thinking Patterns APIThis feature allows users to feed back their chain of thought, enabling the model to maintain a more coherent reasoning process. It is suitable for tasks requiring multi-step reasoning, such as mathematical proofs, programming debugging, and data analysis.
• Sparse attention mechanismDeepSeek-V3.2-Exp introduces DeepSeek Sparse Attention (DSA), which uses the Lightning Indexer to calculate the index score between the query token and the preceding token, selects key value entries, and achieves fine-grained sparse attention, significantly improving the efficiency of long text processing.
• Lightning IndexerThe Lightning Indexer is a core component of DSA. It calculates the index score between the query token and the preceding token, quickly determining which tokens are most important to the query token with a small number of index headers and efficient calculation methods.
• Fine-grained label selectionBased on the index score, the model selects the top k key value entries for attention calculation, reducing unnecessary computation and improving the model's inference speed and efficiency.
• MLA-based implementationDSA is implemented under the Multi-Layer Attention (MLA) architecture and adopts the Multi-Query Attention (MQA) mode, which allows each key value entry to be shared among multiple queries, thus improving computational efficiency.
• Continuous training and optimizationStarting with the basic checkpoints of DeepSeek-V3.1-Terminus, the Lightning Indexer and the entire model are optimized through two phases: intensive warm-up and sparse training, to adapt to the sparse attention pattern.

Project address for DeepSeek-V3.2

• DeepSeek-V3.2https://huggingface.co/deepseek-ai/DeepSeek-V3.2
• DeepSeek-V3.2-Specialehttps://huggingface.co/deepseek-ai/DeepSeek-V3.2-Speciale
• HuggingFace model libraryhttps://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp
• Magic Dash Community: https://modelscope.cn/models/deepseek-ai/DeepSeek-V3.2-Exp
• Technical Papers: https://github.com/deepseek-ai/DeepSeek-V3.2-Exp/blob/main/DeepSeek_V3_2.pdf

How to use DeepSeek-V3.2

• Web version & APP:access DeepSeek Official WebsiteOr download the official app.
• Use via APIDevelopers can use the DeepSeek-V3.2-Exp API to integrate the model's functionality into their applications. The significant reduction in API pricing lowers the cost of use, making it easier for more developers to adopt the technology.
• run locallyUsers can download the model weights for DeepSeek-V3.2-Exp from the Hugging Face platform, follow the provided local running guide to convert the weights to the format required for inference demonstration, and then launch the interactive chat interface to use them.
• Use the official applicationThe official DeepSeek app, web version, and mini-program have all been updated to DeepSeek-V3.2-Exp. Users can directly use the model through these platforms without any additional configuration.
• Model fine-tuningFor specific tasks or domains, users can fine-tune DeepSeek-V3.2-Exp to better adapt to specific application scenarios and improve the model's performance on specific tasks.
• Secondary developmentDeepSeek-V3.2-Exp is open source on the Hugging Face and ModelScope platforms. Users can refer to the open source code and implementation details to understand how the model works and perform secondary development as needed.

Application scenarios of DeepSeek-V3.2

• Long text processingIt is suitable for scenarios that require processing long texts, such as long document analysis and long text generation. Its sparse attention mechanism can significantly improve the processing efficiency of long texts.
• Search and Information RetrievalIt can be used in scenarios such as search agents to help users quickly and accurately retrieve information, improving search efficiency and relevance.
• Code generation and programming assistanceIt supports code generation and programming assistance tasks, such as code completion and code optimization, to help developers improve programming efficiency and code quality.
• Mathematical and Logical ReasoningIt excels in mathematical problem-solving and logical reasoning tasks, and is able to handle complex mathematical problems and logical reasoning chains.
• Multilingual processingIt supports multilingual tasks and can be applied to cross-language text generation, translation and other scenarios to meet the needs of different language environments.
• Intelligent Agents and InteractionsAs the core model of intelligent agents, it can be used to build intelligent assistants, chatbots, etc., and provide natural language interaction services.