# DeepSeek V3.2
Cherry Studio users can now, through the built-in **CherryIN** service, experience for free **DeepSeek V3.2**—DeepSeek’s flagship sparse-attention MoE model released on December 1, 2025, which for the first time natively integrates "thinking" into tool calling, is an ideal choice for advanced agents and long-context scenarios.
***
## 🚀 What is DeepSeek V3.2?
DeepSeek V3.2 is iterated from V3.2-Exp, adopts a Mixture-of-Experts (MoE) architecture, and introduces **DeepSeek Sparse Attention (DSA)** sparse attention mechanism, significantly reducing long-context inference costs while maintaining an ultra-large total parameter scale.
* Architecture: MoE + DeepSeek Sparse Attention (DSA) + Multi-Head Latent Attention (MLA)
* Total parameters: 685B
* Activated parameters per token: about 37B
* Number of experts: 256 experts per layer
* Open-source license: MIT
* Release date: December 1, 2025 (V3.2-Exp released on September 29, 2025)
V3.2 was also released for API as the **DeepSeek-V3.2-Speciale** version, achieving gold-medal-level performance on complex reasoning tasks in IMO, CMO, ICPC World Finals, and IOI 2025.
***
## 📚 Continuing a solid training and alignment process
DeepSeek V3.2 follows the mature training pipeline of the V3 series and makes key extensions for agent scenarios:
1. **Large-scale pretraining**: Base training completed on massive high-quality multilingual corpora, covering code, mathematics, and scientific knowledge.
2. **Introduction of sparse attention**: The main model and lightning indexer were trained at a 128K sequence length, with each query token selecting 2048 key-value tokens to participate in attention.
3. **Large-scale agent data synthesis**: A new agent training data synthesis method covering 1,800+ environments and 85,000+ complex instructions.
4. **Integration of thinking and tool calling**: V3.2 is DeepSeek’s first model to natively integrate "thinking" into tool calling, supporting tool use in both "thinking mode" and "non-thinking mode".
***
## ⚙️ Flagship core capabilities
DeepSeek V3.2 is positioned as having comprehensive capabilities "on par with GPT-5," with major enhancements for agents and complex reasoning:
* ✅ **Native thinking + tool calling**: The first DeepSeek model to integrate thinking into tool use
* ✅ **Top-tier reasoning ability**: V3.2-Speciale reached gold-medal level at IMO / CMO / ICPC World Finals / IOI 2025
* ✅ **Code and development tasks**: Inherits the strong coding capabilities of the V3 series
* ✅ **Long-context stability**: DSA enables analysis of long documents and codebases
* ✅ **Structured tool calling**: Suitable for building agents that do multi-step planning and execution
***
## 💡 DeepSeek Sparse Attention: longer, cheaper
DSA is the core technical upgrade in V3.2, achieved through **lightning indexer + fine-grained token selection** to deliver:
* Fine-grained sparse attention implemented for the first time in a large model
* Reduced the core attention complexity from O(L²)
* Greatly speeds up long-context training and inference while maintaining output quality almost identical to dense attention
| Scenario | Recommended usage | Example |
| --------------------------------- | ----------------------- | ------------------------------------------------------------ |
| Short conversations / simple Q\&A | Direct call | Everyday Q\&A, summarization |
| Moderately complex tasks | Enable tool calling | Data analysis, code refactoring |
| Complex agent tasks | Thinking + tool calling | Multi-step planning, codebase analysis, long document review |
***
## 🌟 Open, usable, ecosystem-friendly
* ⚡ Long-context inference acceleration brought by DSA
* 💰 Through CherryIN in Cherry Studio**use for free**
* 🖥️ Open-source weights, MIT license, day-0 support from mainstream inference frameworks such as vLLM and SGLang
***
## 🧠 Focus on practical capabilities: code and agents
DeepSeek V3.2 performs particularly well in real-world development workflows:
* Multilingual code generation and refactoring
* Repository-level context understanding and patch generation
* Agent toolchain: stable calling of external tools, search, and code execution
* Mathematics and complex reasoning: supports competition-level problems
***
## 🧭 How to use it in Cherry Studio?
1. Open Cherry Studio and go to **Settings → Model Services**.
2. Find **CherryIN** the service provider and enable it.
3. In the model list, select **DeepSeek V3.2**.
4. Return to the chat interface and switch the top model selector to **DeepSeek V3.2** to start chatting.
> 💡 Tip: The free model quota provided by CherryIN is covered by Cherry Studio officially, making it suitable for daily use and evaluation; for production environments, it is recommended to use it together with the official DeepSeek API.
***
📘 **Try DeepSeek V3.2 now and start your journey with flagship-level reasoning and agents!**
---
# Agent Instructions: Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question.
Perform an HTTP GET request on the current page URL with the `ask` query parameter:
```
GET https://docs.cherry-ai.com/docs/en-us/pre-basic/providers/cherryai/mian-fei-ti-yan-deepseek-v3.2-qi-jian-xi-shu-zhu-yi-li-moe.md?ask=
```
The question should be specific, self-contained, and written in natural language.
The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.
Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.