DeepSeek-V3.2 Now on SiliconFlow: Reasoning-first model built for agents

TL;DR: DeepSeek-V3.2 (official version of V3.2-Exp) is now live on SiliconFlow. As a reasoning-first model built for agents, it combines high efficiency with GPT-5-level reasoning performance and a 164K context window. It also features tool-use capabilities in thinking mode, validated across 85K+ complex instructions and 1,800+ environments. Start building today with SiliconFlow's API to supercharge your agentic workflows.

We are thrilled to unlock access to DeepSeek's latest model on SiliconFlow, DeepSeek-V3.2, a new series that harmonizes computational efficiency with superior reasoning and agentic performance. As the first DeepSeek model to integrate thinking directly into tool-use, DeepSeek-V3.2 delivers GPT-5 level reasoning with significantly shorter outputs. Meanwhile, DeepSeek-V3.2-Speciale pushes open-source boundaries of theorem proving and coding to rival Gemini 3 Pro. Together, they set a new benchmark for developers building next-generation AI agents.

Now, through SiliconFlow's DeepSeek-V3.2 API, you can expect:

• Cost-effective Pricing:

  • DeepSeek-V3.2 $0.27/M tokens (input) and $0.42/M tokens (output)

  • DeepSeek-V3.2-Speciale is coming soon and stay tuned for first-hand updates

• 164K Context Window: Perfect for long documents, complex multi-turn conversations, and extended agentic tasks.

• Seamless Integration: Instantly deploy via SiliconFlow's OpenAI-compatible API, or plug into your existing stack through Claude Code, Gen-CLI and Cline.

  
  

Whether you're building agents, coding assistants, or complex reasoning pipelines, SiliconFlow's DeepSeek-V3.2 API delivers the performance you need at a fraction of the expected cost and latency.

Why it matters

For developers building agents, multi-step reasoning pipelines or any AI system that needs to think and act, the DeepSeek-V3.2 series finally delivers the combination the industry has been waiting for: frontier-grade reasoning, integrated tool-use during thinking, and real-world efficiency:

1. World-Leading Reasoning Capabilities

• DeepSeek-V3.2: The Efficient "Daily Driver" for Agents

Engineered to strike the perfect balance between reasoning capabilities and output length, DeepSeek-V3.2 is your go-to choice for production workflows, such as advanced Q&A and general agent tasks.

• Performance: Delivers reasoning capabilities on par with GPT-5.

• Efficiency: Compared to Kimi-K2-Thinking, V3.2 has significantly shorter output lengths, translating to lower computational overhead and reduced overall generation time.

• DeepSeek-V3.2-Speciale: Maxed-out reasoning capabilities (Research Preview)

As the enhanced long-thinking variant of V3.2, V3.2-Speciale aims to push the boundaries of open-source reasoning capabilities, integrating the theorem-proving capabilities of DeepSeek-Math-V2.

• Gold-Medal Performance: V3.2-Speciale attains gold-level results in IMO, CMO, ICPC World Finals & IOI 2025.

• Benchmarks: It excels in complex instruction following, rigorous mathematical reasoning and logical verification, effectively rivaling Gemini 3 Pro on mainstream reasoning leaderboards.

  
  

The number in parentheses indicates the approximate total token consumption.

2. Thinking in Tool-Use

DeepSeek-V3.2 breaks the barrier between "reasoning" and "acting." Unlike previous versions where tool usage was restricted during the thinking process, DeepSeek-V3.2 is the first to seamlessly integrate thinking directly into tool-use, supporting tool invocation in both Thinking and Non-Thinking modes.

To deliver this level of agentic reliability, DeepSeek introduces a massive-scale training synthesis method:

• Robust Generalization: The model was forged through "hard-to-solve, easy-to-verify" reinforcement learning tasks.

• Extensive Coverage: Training spanned 1,800+ distinct environments and over 85,000+ complex instructions, significantly enhancing the model's generalization and instruction-following capability in the agent context.

  
  

What this means for your workflows:

DeepSeek-V3.2 emerges as a highly cost-efficient alternative in agent scenarios, significantly narrowing the performance gap between open and frontier proprietary models while incurring substantially lower costs — all available via SiliconFlow's API.

Here's an example built using SiliconFlow's DeepSeek-V3.2 API, demonstrating how the model can assist in writing, optimizing, and reasoning about real-time interactive code.

What makes it powerful

DeepSeek-V3.2 series' performance is enabled by three core technical breakthroughs:

• DeepSeek Sparse Attention (DSA):

To tackle the challenge of long-context processing, the model introduces DeepSeek Sparse Attention (DSA). This efficient attention mechanism substantially reduces computational complexity without compromising performance, specifically optimized for long-context scenarios.

• Scalable Reinforcement Learning:

DeepSeek-V3.2 leverages a robust Reinforcement Learning (RL) protocol combined with scaled post-training compute. This advanced training framework is the key driver behind the model's exceptional reasoning capabilities.

• Large-Scale Agentic Task Synthesis Pipeline:

DeepSeek has revolutionized agent capability through a novel Large-Scale Agentic Task Synthesis Pipeline. By systematically generating training data at scale, the model integrates reasoning directly into tool-use scenarios. This results in superior compliance and generalization, ensuring that your agents can reliably navigate complex, multi-step interactive environments with precision.

Developer-Ready Integration

Beyond DeepSeek-V3.2's industry-leading agentic performance, SiliconFlow delivers instant compatibility with your existing development ecosystem:

• OpenAI-Compatible Tools: Seamless integration with Cline, Qwen Code, Gen-CLI, and other standard development environments — just plug in your SiliconFlow API key.

• Anthropic-Compatible API: Works with Claude Code and any Anthropic-compatible tools for code reviews, debugging, and architectural refactoring.

• Platform Integrations: Ready-to-use in Dify, ChatHub, Chatbox, Sider, MindSearch, DB-GPT, and also available through OpenRouter.

With powerful models, seamless integrations, and competitive pricing, SiliconFlow transforms how you build — letting you ship faster and scale smarter.

Get Started Immediately

1. Explore: Try DeepSeek-V3.2 in the SiliconFlow playground.

2. Integrate: Use our OpenAI-compatible API. Explore the full API specifications in the SiliconFlow API documentation.

      import requests

url = "https://api.siliconflow.com/v1/chat/completions"

payload = {
    "model": "deepseek-ai/DeepSeek-V3.2",
    "messages": [
        {
            "role": "user",
            "content": "an island near sea, with seagulls, moon shining over the sea, light house, boats int he background, fish flying over the sea"
        }
    ],
    "stream": True,
    "max_tokens": 4096,
    "enable_thinking": False,
    "thinking_budget": 4096,
    "min_p": 0,
    "stop": "1",
    "temperature": 0.7,
    "top_p": 0.7,
    "top_k": 50,
    "frequency_penalty": 0.5,
    "n": 1,
    "response_format": { "type": "json_object" },
    "tools": [
        {
            "type": "function",
            "function": {
                "name": "<string>",
                "description": "<string>",
                "parameters": {},
                "strict": False
            }
        }
    ]
}
headers = {
    "Authorization": "Bearer <token>",
    "Content-Type": "application/json"
}

response = requests.post(url, json=payload, headers=headers)

print(response.text)

• Business or Sales Inquiries →

• Join our Discord community now →

• Follow us on X for the latest updates →

• Explore all available models on SiliconFlow →