DeepSeek-V3.2-Exp Now on SiliconFlow: Efficient Long-Context Reasoning at 50% Lower Cost

TL;DR: DeepSeek-V3.2-Exp, the latest experimental model in the DeepSeek series, is now live on SiliconFlow. Featuring DeepSeek Sparse Attention (DSA) for fine-grained efficiency, it delivers faster and better long-context reasoning, coding, and agent search performance — with dense-level quality, 164K context window, and seamless integration via SiliconFlow's product-ready API service.

We're excited to announce that DeepSeek-V3.2-Exp, the latest experimental model in the DeepSeek series, is now available on SiliconFlow. This release offers a glimpse into the future of efficient long-context reasoning. Built as a bridge toward DeepSeek's next-generation architecture, DeepSeek-V3.2-Exp introduces DeepSeek Sparse Attention (DSA) — a fine-grained mechanism that significantly enhances training and inference efficiency across extended text, code, and reasoning tasks.

With SiliconFlow API, you can expect:

• Cost-effective Pricing: DeepSeek-V3.2-Exp $0.27/M tokens (input) and $0.41/M tokens (output).

• Longer Contexts, Less Overhead: 164K context window allows you to analyze or generate longer documents, conversations, or scripts without costly truncation.

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

With the powerful model, seamless integrations, and cost-effective pricing, SiliconFlow transforms how developers build — letting you ship faster and scale smarter.

Why This Release Matters

DeepSeek-V3.2-Exp builds upon DeepSeek-V3.1-Terminus, marking a major step in DeepSeek's research toward more efficient transformer architectures.

Its new DeepSeek Sparse Attention mechanism achieves fine-grained sparsity for the first time, delivering notable speed and memory gains in long-context training and inference — while preserving the same output quality as dense attention.

To ensure rigorous evaluation, DeepSeek-V3.2-Exp was trained under the same setup as DeepSeek-V3.1-Terminus, and across multiple public benchmarks it demonstrates comparable performance and stability, validating sparse attention as a reliable path toward greater efficiency.

Access SiliconFlow's DeepSeek-V3.2-Exp API Service

Let's start by checking out DeepSeek-V3.2-Exp's performance in the SiliconFlow Playground.

When tested with a classic "two-train" problem, DeepSeek-V3.2-Exp reasoned through the steps efficiently—identifying key variables, building equations, and verifying the result. Most importantly, the final answer — 171 ¹⁄₉ km — was correct and clearly reasoned.

Besides the interactive playground, SiliconFlow also delivers instant compatibility with your existing development ecosystem:

• Anthropic Ecosystem Support: Full compatibility with Anthropic's ecosystem for comprehensive code reviews, bug tracing, and architectural refactoring using DeepSeek-V3.2-Exp in Claude Code.

• Developer Tool Integration: Seamless compatibility with Cline, Qwen Code, Gen-CLI, and other OpenAI-standard development tools — simply use your SiliconFlow API to access DeepSeek-V3.2-Exp's capabilities.

• Other Supported Platforms: Ready-to-use integrations with Dify, Sider, MindSearch, DB-GPT, ChatHub, and Chatbox, leveraging DeepSeek-V3.2-Exp and other powerful models.

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

Get Started Immediately

Ready to dive deeper? Here's how to get started:

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

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

import requestsurl = "https://api.siliconflow.com/v1/chat/completions"payload = {    "model": "deepseek-ai/DeepSeek-V3.2-Exp",    "messages": [        {            "role": "user",            "content": "Tell me a story"        }    ],    "stream": False,    "max_tokens": 4096,    "enable_thinking": False,    "thinking_budget": 4096,    "min_p": 0.05,    "stop": None,    "temperature": 0.7,    "top_p": 0.7,    "top_k": 50,    "frequency_penalty": 0.5,    "n": 1,    "response_format": { "type": "text" },    "tools": [        {            "type": "function",            "function": {                "description": "<string>",                "name": "<string>",                "parameters": {},                "strict": False            }        }    ]}headers = {    "Authorization": "Bearer <token>",    "Content-Type": "application/json"}response = requests.post(url, json=payload, headers=headers)print(response.json())

Start building with DeepSeek-V3.2-Exp on SiliconFlow today — faster, smarter, and more cost-effective AI for your applications.

Business or Sales Inquiries →

Join our Discord community now →

Follow us on X for the latest updates →

Explore all available models on SiliconFlow →