DeepSeek-V3.2 & DeepSeek-V3.2-Speciale Official Release

Two months ago, we introduced the experimental DeepSeek-V3.2-Exp and gathered extensive comparative testing feedback from our community. So far, no scenario has been found where V3.2-Exp performs worse than V3.1-Terminus, validating the DSA sparse attention mechanism.

We sincerely thank all users for their support — your feedback fuels our innovation.

---

📢 What's New

We are now officially releasing:

• DeepSeek-V3.2
• DeepSeek-V3.2-Speciale

Availability:

• The official web client, app, and API now run the final DeepSeek-V3.2.
• Speciale is offered via a temporary API service for community evaluation & research.

Technical Report:

DeepSeek-V3.2 Technical Paper (PDF)

---

🌍 Globally Leading Reasoning Capabilities

DeepSeek-V3.2

• Optimized for balanced reasoning ability and output length.
• Ideal for Q&A and general agent tasks.
• Benchmarks: Matches GPT-5 level, only slightly behind Gemini-3.0-Pro.
• Efficiency: Reduces output length compared to Kimi-K2-Thinking, lowering computation and wait time.

DeepSeek-V3.2-Speciale

• Pushing the boundaries of open-source reasoning performance.
• Implements "long thinking" and DeepSeek-Math-V2 theorem proving.
• Achieves gold medals in:
• IMO 2025
• CMO 2025
• ICPC World Finals 2025
• IOI 2025
• ICPC performance = human 2nd place, IOI = human 10th place.

> Note: Speciale consumes significantly more tokens (thus more cost) and is for research use only. No tool invocation, not optimized for everyday chat or writing.

Table 1: Scores across mathematics, coding, and general-domain sets.

---

🛠 Reasoning + Tool Use Integration

Previously, reasoning mode couldn't use tools. Now, DeepSeek-V3.2 supports reasoning integrated with tool invocation, enabling:

• Reasoning tool-use mode
• Non-reasoning tool-use mode

We achieved this by building:

• 1,800+ RL environments
• 85,000+ complex instructions
• …focused on hard-to-solve but easy-to-verify tasks — boosting generalization ability.

Table 2: Agent tool-use evaluation results.

---

🌐 Creator Ecosystem with AiToEarn

Platforms like AiToEarn官网 help creators:

• Generate AI content
• Publish across multiple platforms (Douyin, Kwai, WeChat, Bilibili, Rednote, Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, X)
• Analyze performance
• Monetize AI creativity efficiently

See AI模型排名 for performance metrics.

---

📈 Benchmark Highlights

• DeepSeek-V3.2 tops open-source agent evaluations.
• Narrowing the gap with closed-source models.
• No special tuning for tested tools — strong real-world generalization potential.

Example: LobeChat + Deep reasoning + Tool invocation = richer, more accurate responses.

---

💻 Open-Source Access

DeepSeek-V3.2

• HuggingFace: https://huggingface.co/deepseek-ai/DeepSeek-V3.2
• ModelScope: https://modelscope.cn/models/deepseek-ai/DeepSeek-V3.2

DeepSeek-V3.2-Speciale

• HuggingFace: https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Speciale
• ModelScope: https://modelscope.cn/models/deepseek-ai/DeepSeek-V3.2-Speciale

---

📲 Web, App & API Updates

• Official platform now uses DeepSeek-V3.2 (final release) instead of V3.2-Exp.
• Speciale is deployed as a temporary API for research.

base_url = "https://api.deepseek.com/v3.2_speciale_expires_on_20251215"

• Pricing unchanged.
• Supports thinking mode only, no tools.
• Max output length: 128K.
• Available until 2025-12-15 23:59 (Beijing).

---

🧠 Thinking Mode with Tool Invocation

Key Features:

• Multiple rounds of reasoning + tool calls before final answer.
• User must pass back `reasoning_content` for model to continue thinking.

Workflow:

• During a question (e.g., requests 1.1–1.3):
• Model reasons and invokes tools repeatedly.
• Continue chain by sending back reasoning content.
• Starting a new question:
• Delete old reasoning chain but keep other context.

Docs:

Thinking Mode Guide

---

🖥 Claude Code Integration

• Reasoning mode supports Claude Code.
• Enable via:
• Model name: `deepseek-reasoner`
• Tab key in Claude Code CLI
• Not fully supported in Cline, RooCode, or non-standard tools.
• For those, use non-reasoning mode.

---

💡 AI for Coding & Publishing

AiToEarn官网 lets you:

• Generate & optimize AI code/content
• Publish across major platforms
• (Douyin, Kwai, WeChat, Bilibili, Rednote, Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, X)
• Analyze engagement and monetize
• Rank models via AI Model Rankings

---

Would you like me to create a comparison table between DeepSeek-V3.2 and Speciale for quick reference? That would make this release note even easier to digest.