DeepSeek V3.2

DeepSeek V3.2 Official Release: Enhanced Agent Capabilities with Integrated Reasoning

Two months ago, we released the experimental DeepSeek-V3.2-Exp and received numerous comparison test results from enthusiastic users. So far, we have not found V3.2-Exp to be significantly worse than V3.1-Terminus in any specific scenario, which validates the effectiveness of the DSA sparse attention mechanism. We also thank all users for their continuous positive feedback and support, which has injected more confidence and momentum into our ongoing innovation.

Today, we are releasing two official models simultaneously: DeepSeek-V3.2 and DeepSeek-V3.2-Speciale. The official web interface, App, and API have all been updated to the official DeepSeek-V3.2, and we welcome you to use it. The Speciale version is currently available only as a temporary API service for community evaluation and research.

The technical report for the new models has been released simultaneously:DeepSeek-V3.2 Technical Report - ModelScope

• DeepSeek-V3.2 aims to balance reasoning capability and output length, making it suitable for daily use, such as Q&A scenarios and general agent tasks. In public reasoning benchmarks, DeepSeek-V3.2 achieves the level of GPT-5, only slightly below Gemini-3.0-Pro; compared to Kimi-K2-Thinking, V3.2 significantly reduces output length, greatly decreasing computational overhead and user waiting time.

• DeepSeek-V3.2-Speciale aims to push the reasoning capability of open-source models to the extreme, exploring the boundaries of model capabilities. V3.2-Speciale is a long-thinking enhanced version of DeepSeek-V3.2, combined with the theorem-proving capabilities of DeepSeek-Math-V2. This model exhibits excellent instruction following, rigorous mathematical proof, and logical verification capabilities, with performance comparable to Gemini-3.0-Pro on mainstream reasoning benchmarks (see table below). More notably, the V3.2-Speciale model has successfully won gold medals at IMO 2025 (International Mathematical Olympiad), CMO 2025 (Chinese Mathematical Olympiad), ICPC World Finals 2025 (International Collegiate Programming Contest World Finals), and IOI 2025 (International Olympiad in Informatics). Among these, the ICPC and IOI results reached the level of the second and tenth place human contestants, respectively.

Tips: On highly complex tasks, the Speciale model significantly outperforms the standard version, but it also consumes significantly more tokens and incurs higher costs. Currently, DeepSeek-V3.2-Speciale is for research use only, does not support tool calling, and has not been specifically optimized for daily conversation and writing tasks.

• Unlike previous versions that could not call tools in thinking mode, DeepSeek-V3.2 is the first model we have released that integrates thinking into tool use, and it supports tool calling in both thinking and non-thinking modes. We have proposed a large-scale agent training data synthesis method, constructing a large number of reinforcement learning tasks that are "hard to answer but easy to verify" (1,800+ environments, 85,000+ complex instructions), greatly improving the model's generalization capability.

As shown in the table above, the DeepSeek-V3.2 model achieves the highest level among current open-source models in agent evaluations, significantly narrowing the gap between open-source and closed-source models. It is worth noting that V3.2 was not specifically trained on the tools in these test sets, so we believe V3.2 can demonstrate strong generalization in real-world application scenarios.

• DeepSeek-V3.2

HuggingFace:DeepSeek-V3.2

• ModelScope:DeepSeek-V3.2

• DeepSeek-V3.2-Speciale

HuggingFace:DeepSeek-V3.2-Speciale

• ModelScope:DeepSeek-V3.2-Speciale

DeepSeek-V3.2 is our current officially served model. The official website, APP, and API models have all been upgraded from DeepSeek-V3.2-Exp to the official DeepSeek-V3.2, with no change in usage.

Meanwhile, to facilitate community evaluation and research, we have informally deployed an API service for DeepSeek-V3.2-Speciale. API users can access this model by setting base_url="https://api.deepseek.com/v3.2_speciale_expires_on_20251215". The API price for this model remains unchanged, it only supports conversation in thinking mode, does not support tool calling, and the default maximum output length is 128K. The service is available until 23:59 Beijing time on 2025-12-15.

This API update supports tool calling in thinking mode for DeepSeek-V3.2. Currently, in thinking mode, the model can undergo multiple rounds of thinking and tool calling to ultimately provide more detailed and accurate answers.

• During the process of answering question 1 (requests 1.1 - 1.3), the model performs multiple thinking and tool calls before giving an answer. In this process, the user needs to return the chain-of-thought content (reasoning_content) to the API to allow the model to continue thinking.

• At the start of the next user question (request 2.1), the previous chain-of-thought should be deleted, and other content should be retained when sending to the API.

• For more detailed usage, please refer to the API documentation:Thinking Mode Guide - API Documentation

DeepSeek-V3.2's thinking mode also adds support for Claude Code. Users can use it by changing the model name to deepseek-reasoner, or by pressing the Tab key in the Claude Code CLI to enable thinking mode. However, it should be noted that thinking mode is not fully adapted to components that use non-standard tool calling, such as Cline and RooCode. We recommend that users continue to use non-thinking mode when using such components.