DeepSeek 发布 V3.2 与 V3.2-Speciale：面向智能体的推理模型

DeepSeek API Docs Logo DeepSeek API Docs Logo # DeepSeek-V3.2 Release 🚀 Launching DeepSeek-V3.2 & DeepSeek-V3.2-Speciale — Reasoning-first models built for agents! 🔹 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 report: # 🏆 World-Leading Reasoning 🔹 V3.2: Balanced inference vs. length. Your daily driver at GPT-5 level performance. 🔹 V3.2-Speciale: Maxed-out reasoning capabilities. Rivals Gemini-3.0-Pro. 🥇 Gold-Medal Performance: V3.2-Speciale attains gold-level results in IMO, CMO, ICPC World Finals & IOI 2025. [...] 📝 Note: V3.2-Speciale dominates complex tasks but requires higher token usage. Currently API-only (no tool-use) to support community evaluation & research. # 🤖 Thinking in Tool-Use 🔹 Introduces a new massive agent training data synthesis method covering 1,800+ environments & 85k+ complex instructions. 🔹 DeepSeek-V3.2 is

##### Mixed RL Training For DeepSeek-V3.2, we still adopt Group Relative Policy Optimization (GRPO) (deepseekmath; deepseekr1) as the RL training algorithm. As DeepSeek-V3.2-Exp, we merge reasoning, agent, and human alignment training into one RL stage. This approach effectively balances performance across diverse domains while circumventing the catastrophic forgetting issues commonly associated with multi-stage training paradigms. For reasoning and agent tasks, we employ rule-based outcome reward, length penalty, and language consistency reward. For general tasks, we employ a generative reward model where each prompt has its own rubrics for evaluation. ##### DeepSeek-V3.2 and DeepSeek-V3.2-Speciale [...] Notably, with the aim of pushing the boundaries of open models in the reasoning domain, we relaxed the length constraints to develop DeepSeek-V3.2-Speciale. As a result, DeepSeek-V3.2-Speciale achieves performance parity with the leading closed-source system, Gemini-3.0-Pro (gemini3

No, Anakin said. There is another. DeepSeek: 🚀 Launching DeepSeek-V3.2 & DeepSeek-V3.2-Speciale — Reasoning-first models built for agents! 🔹 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 report [here]]( v3.2 model, v3.2-speciale model. 🏆 World-Leading Reasoning 🔹 V3.2: Balanced inference vs. length. Your daily driver at GPT-5 level performance. 🔹 V3.2-Speciale: Maxed-out reasoning capabilities. Rivals Gemini-3.0-Pro. 🥇 Gold-Medal Performance: V3.2-Speciale attains gold-level results in IMO, CMO, ICPC World Finals & IOI 2025. [...] Teortaxes: V3.2 is here, it’s no longer “exp”. It’s frontier. Except coding/agentic things that are being neurotically benchmaxxed by the big 3. That’ll take one more update. “Speciale” is a high compute variant that’s between Gemini and GPT-5 and can score gold on IMO-2025. Thank you guys. hallerite: hmm, I wond

DeepSeek-V3.2: Designed to balance strong reasoning with shorter, more efficient outputs. It is suitable for everyday use, including Q&A and general agent tasks. DeepSeek-V3.2-Speciale (high-compute variant): Created to push open-source reasoning to the limit. It’s an enhanced long-thinking version of V3.2, further strengthened with the theorem-proving abilities of DeepSeek-Math-V2. According to the research paper, DeepSeek acknowledges that open-source models still lag behind the best proprietary models. In complex tasks, closed-source models have been improving faster, widening the performance gap. They identified three core issues holding open-source models back: [...] And just a week later, DeepSeek open-sourced DeepSeek-V3.2-Exp, which builds on V3.1-Terminus with the introduction of DeepSeek Sparse Attention, a mechanism to optimize training and inference efficiency in long-context scenarios. Across public benchmarks in various domains, V3.2-Exp demonstrates performance on p

specific aircraft we're using. As we saw in the beginning, it did. [laughter] Deepseek has released a new model, V3.2, which is designed to be the official successor to V3.2-EXP or experimental, which did come out a little over 2 months ago. However, something of interest is that they also released V3.2-p 2-p special which is something I have only heard reserved for like special Italian sports cars but now DeepS has a special of their own and that is designed to push the boundaries of reasoning capabilities. It is currently API only and seems to be in some form of kind of a research beta period where folks will be able to access it. They do have these specific kind of updates or instructions on how to access that via a temporary endpoint right here that will be good until December 15th. [...] AI Integration & Consulting: Join the Discord: In this video, we take a look at the newly released DeepSeek V3.2 model from DeepSeek. This version is the official release of the earlier V3.2-EX

user avatar DeepSeek @deepseek\_ai Dec 1, 2025 🛠 Open Source Release 📦 DeepSeek-V3.2 Model: huggingface.co/deepseek-ai/De… 📦 DeepSeek-V3.2-Speciale Model: huggingface.co/deepseek-ai/De… 📄 Tech report: huggingface.co/deepseek-ai/De… 5/n deepseek-ai/DeepSeek-V3.2 · Hugging FaceFrom huggingface.co 306K user avatar Z.ai @Zai\_org Dec 1, 2025 Legend❤️ 121K # Join the conversation Read 951 more replies [...] user avatar DeepSeek @deepseek\_ai Dec 1, 2025 🤖 Thinking in Tool-Use 🔹 Introduces a new massive agent training data synthesis method covering 1,800+ environments & 85k+ complex instructions. 🔹 DeepSeek-V3.2 is our first model to integrate thinking directly into tool-use, and also supports tool-use in both thinking and 216K user avatar DeepSeek @deepseek\_ai Dec 1, 2025 💻 API Update 🔹 V3.2: Same usage pattern as V3.2-Exp. 🔹 V3.2-Speciale: Served via a temporary endpoint: base\_url="api.deepseek.com/v3.2