Overview

DeepSeek V3.2 Exp specialises in Chinese-language tasks and mathematical reasoning. It balances accuracy with low token costs, making it ideal for education platforms, community moderation, and support assistants.

Sample Request

curl -X POST "https://api.transendai.net/v1/chat/deepseek-v3.2" \
  -H "Authorization: Bearer $TRANSEND_API_KEY" \
  -H "Content-Type": "application/json" \
  -d '{
    "messages": [
      { "role": "system", "content": "你是一名教学助手，善于归纳总结和出题。" },
      { "role": "user", "content": "请把这篇文章总结为 5 条要点，并给出一道思考题。" }
    ],
    "reasoning": { "chain_of_thought": "retain" },
    "response_format": { "type": "markdown" }
  }'

chain_of_thought can be retain, summarise, or hide.

Tool Calling

"tools": [
  {
    "name": "getLectureNotes",
    "description": "Fetch lecture notes by ID",
    "parameters": {
      "type": "object",
      "properties": { "lecture_id": { "type": "string" } },
      "required": ["lecture_id"]
    }
  }
]

When the assistant calls a tool, execute it and append the result as a new assistant message with role: "tool".

Response Example

{
  "choices": [
    {
      "message": {
        "role": "assistant",
        "content": "### 要点\n1. 模型采用混合专家结构...\n\n**思考题**：..."
      }
    }
  ],
  "usage": { "prompt_tokens": 2100, "completion_tokens": 620 }
}

Error Codes

Code	Meaning	Resolution
`400`	Unsupported reasoning mode.	Use `retain`, `summarise`, or `hide`.
`403`	Regional restriction.	Ensure the workspace has access to DeepSeek models.
`429`	Token or RPM limit exceeded.	Implement retries with exponential backoff.

Optimisation Tips

Keep prompts concise; Chinese tokens consume fewer credits, but large contexts still add up.
Use response_format: { "type": "json_schema" } when integrating with grading systems.
Record trace_id for analytics and hit correlation in the observability dashboard.