DeepSeek-V3.2 and DeepSeek-V3.2-Speciale Release

DeepSeek has launched DeepSeek-V3.2 and DeepSeek-V3.2-Speciale, a new series of reasoning-first models designed specifically for agentic workflows. These models introduce the ability to integrate thinking processes directly into tool-use, significantly enhancing the model's ability to handle complex instructions across diverse environments.

World-Leading Reasoning Capabilities

DeepSeek-V3.2 and V3.2-Speciale offer different optimization paths for reasoning tasks:

  • DeepSeek-V3.2: Positioned as a balanced model for daily use, providing performance comparable to GPT-5 level performance.
  • DeepSeek-V3.2-Speciale: A model optimized for maximum reasoning capabilities, rivaling Gemini-3.0-Pro. It has achieved gold-medal level results in the IMO, CMO, ICPC World Finals, and IOI 2025.

While V3.2-Speciale dominates complex tasks, it requires higher token usage and is currently available exclusively via API for community evaluation and research purposes.

Thinking in Tool-Use and Agent Training

DeepSeek-V3.2 is the first model in the series to integrate "thinking" directly into tool-use. This allows the model to support tool-use in both thinking and non-thinking modes.

This capability is powered by a new massive agent training data synthesis method. The training data covers over 1,800 environments and more than 85,000 complex instructions, enabling the model to act as a more effective agent in complex digital environments.

API Access and Availability

Access to the new models is available through the following channels:

  • DeepSeek-V3.2: Available on the App, Web, and API, following the same usage patterns as the previous V3.2-Exp version.
  • DeepSeek-V3.2-Speciale: Provided via a temporary API endpoint (https://api.deepseek.com/v3.2_speciale_expires_on_20251215). This model is available until December 15th, 2025, at 15:59 UTC. It shares the same pricing as V3.2 but does not support tool calls.

Open Source Release

Both models have been released as open source on Hugging Face:

Technical details regarding the architecture and training are available in the official tech report.

Sources

Related

  • Dispatch
    DeepSeek-V3.1 Release Notes

    DeepSeek-V3.1 introduces hybrid inference modes, improved agentic capabilities, and continued pretraining for long context extension, marking a step toward agent-centric AI.

  • Dispatch
    DeepSeek-V4-Pro GA Release

    DeepSeek has released DeepSeek-V4-Pro, featuring enhanced agent capabilities, flexible reasoning effort settings, and native OpenAI Responses API support.

  • Dispatch
    DeepSeek-V3-0324 Release Notes

    DeepSeek has released DeepSeek-V3-0324, featuring improved reasoning, front-end development, and tool-use capabilities, now available under the MIT License.

  • Dispatch
    DeepSeek-V3-0324 Release Notes

    DeepSeek has released DeepSeek-V3-0324, an updated version of the base model for R1 that features improved instruction following, coding, and math capabilities under an MIT license.

  • Dispatch
    DeepSeek-V4 Preview Release

    DeepSeek has released the open-source DeepSeek-V4 preview, introducing two model variants—Pro and Flash—featuring a 1M token context window and state-of-the-art agentic coding capabilities.