DeepSeek-V3.2 released: Efficient reasoning and general-purpose agent capabilities with sparse attention

1. Abstract

DeepSeek-V3.2 is an officially released version based on V3.2-Exp, focusing on optimizing inference efficiency and output length, and using the DSA sparse attention mechanism to improve long-context performance. DeepSeek-V3.2-Speciale focuses on extreme mathematical reasoning, programming competitions, and rigorous logic verification, and has performed well in many international competitions. Currently, the web side, APP, and API have all been upgraded to DeepSeek-V3.2, and Speciale is open to the community in the form of a temporary API.

2. Core Features

1. Sparse attention and efficient long context

1. DSA sparse attention improves the efficiency of long text tasks, and there is no degradation compared to V3.1-Terminus in community testing.
2. The official version of the model pays more attention to the balance of "reasoning ability-output length", which is more suitable for daily use scenarios.

1. Reasoning ability and competition

1. performance DeepSeek-V3.2 reached the level of GPT-5 in the public reasoning benchmark, only slightly lower than Gemini-3.0-Pro.
2. V3.2-Speciale is an enhanced version of Long-Thinking, combining the proof capabilities of DeepSeek-Math-V2 to achieve gold medal levels in IMO, CMO, ICPC, and IOI 2025.

1. Thinking mode + tool call

1. V3.2 supports combining tool calls in thinking mode for the first time, and you can choose the appropriate tool to complete complex tasks after multiple rounds of thinking.
2. The massive reinforcement learning data that is difficult to solve but easy to verify significantly improves the generalization ability of the model's agents.

3. Installation

1. Get the model

1. and download DeepSeek-V3.2 and Speciale weights through HuggingFace or ModelScope.
2. Speciale is currently for research use only and is not recommended for general build tasks.

2.

1. is recommended to use vLLM or Transformers in the inference environment, combined with FP8 or tensor to improve performance.
2. If you are dealing with very long contexts, you should configure the maximum sequence length according to the hardware resources.

4. Typical use cases

1. DeepSeek-V3.2 (official version)

1. daily conversation, knowledge Q&A and text generation.
2. Agent toolchain tasks, including searches, database queries, and API calls.
3. Office process automation, such as table processing, document generation, and script assistance.

1. DeepSeek-V3.2-Speciale (Research Edition)

1. Difficult mathematical proof and complex reasoning tasks.
2. Algorithm design and logic verification of programming competition scenarios.
3. Explainable reasoning and long-chain reasoning experiments in academic research fields.

5. Ecology and Competing Products

1. The ecological

1. web page, mobile terminal and API have been fully upgraded to DeepSeek-V3.2.
2. Thinking patterns are available in Claude Code via deepseek-reasoner, but some external toolchains are not yet fully adapted.

1. Comparison of competing products

1. Compared with GPT-5 and Gemini-3.0-Pro, it is close to closed-source advanced models in terms of reasoning ability, but it is open-source community-friendly.
2. Compared with open source models (such as Qwen and Llama series), it has stronger generalization in tool call evaluation.

6. Limitations and Precautions

1. The Speciale has a longer output, higher token consumption, and significantly higher operating costs than the Standard Edition.
2. Speciale's temporary API only supports thinking mode and does not support tool calls, making it suitable for evaluation and research.
3. Thinking patterns need to be properly managed reasoning_content, otherwise redundant reasoning or additional overhead may be incurred.
4. In high-risk applications, manual review is still required to avoid wrong decisions caused by model reasoning bias.

7. Project address

https://huggingface.co/deepseek-ai/DeepSeek-V3.2
https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Speciale
https://modelscope.cn/models/deepseek-ai/DeepSeek-V3.2
https://modelscope.cn/models/deepseek-ai/DeepSeek-V3.2-Speciale

8. FAQs

Q: How is the positioning of DeepSeek-V3.2 different from Speciale?

A: V3.2 is for daily use and Agent scenarios, and Speciale is for extreme reasoning and event-level reasoning tasks.

Q: What are the limitations of Speciale's temporary API?

A: It only supports thinking mode, does not support tool calling, and has a limited service period, making it not suitable as a long-term production model.

Q: How to use the thinking mode correctly?

A: Each round of requests needs to be sent back reasoning_content to clear the chain of thought before starting a new issue to avoid logical confusion.

Q: Is it suitable for on-premises deployment?

A: Localized deployment is possible, but GPU and memory resources need to be evaluated based on long-context requirements.