DeepSeek just released DeepSeek-V3.2 and DeepSeek-V3.2-Speciale—reasoning-first open-weight models built for agents. V3.2 is the official successor to V3.2-Exp and is now live on App, Web, and API. The Speciale variant pushes reasoning capabilities to the limit.
🚀 Launching DeepSeek-V3.2 & DeepSeek-V3.2-Speciale — Reasoning-first models built for agents!
— DeepSeek (@deepseek_ai) December 1, 2025
🔹 DeepSeek-V3.2: Official successor to V3.2-Exp. Now live on App, Web & API.
🔹 DeepSeek-V3.2-Speciale: Pushing the boundaries of reasoning capabilities. API-only for now.
📄 Tech… pic.twitter.com/SC49UdmCZv
World-leading reasoning#
DeepSeek-V3.2 is positioned as your daily driver at GPT-5 level performance, balancing inference speed vs. response length.
DeepSeek-V3.2-Speciale maxes out reasoning capabilities, rivaling Gemini-3.0-Pro. It achieves gold-medal results in the 2025 IMO, CMO, ICPC World Finals, and IOI—the same competitive programming and math olympiads that define frontier AI capability.
Key benchmark results for V3.2-Speciale:
- 96.0% on AIME 2025 (vs GPT-5 High’s 94.6%)
- 99.2% on HMMT Feb 2025 (vs Gemini-3.0-Pro’s 97.5%)
- 2701 Codeforces rating (77k tokens)—notably, GPT-5 High edges ahead here at 2708
Thinking in tool-use#
This is DeepSeek’s first model to integrate thinking directly into tool-use. V3.2 supports tool-use in both thinking and non-thinking modes—a capability gap that has limited open models in agentic workflows.
Behind this: a new massive agent training data synthesis method covering 1,800+ environments and 85,000+ complex instructions. This systematic approach to agentic post-training improves generalization and instruction-following in real-world tool-use scenarios.
API and availability#
- V3.2: Same usage pattern as V3.2-Exp, full tool-use support
- V3.2-Speciale: Temporary endpoint at
api.deepseek.com/v3.2_speciale_..., same pricing as V3.2, no tool calls, available until December 15, 2025
Note: V3.2-Speciale dominates complex tasks but requires higher token usage. It’s currently API-only without tool-use to support community evaluation and research.
Open source release#
Both models and the full technical report are available on Hugging Face:
Why this matters#
The gap between open and closed-source AI has been widening. DeepSeek-V3.2 directly challenges that trend—demonstrating that open models can compete with GPT-5 on daily tasks and approach Gemini-3.0-Pro on complex reasoning.
For developers building AI agents, this means frontier-level reasoning and tool-use capabilities without vendor lock-in.