Gemini 3.1 Flash-Lite 預先發布版

這是我們成本效益最高的多模態模型，可為執行頻率高、輕量的工作提供最快的效能。Gemini 3.1 Flash-Lite 最適合處理大量代理工作、簡單的資料擷取作業，以及預算和速度是主要限制的極低延遲應用程式。

gemini-3.1-flash-lite-preview

開發人員指南

Gemini 3.1 Flash-Lite 最擅長處理大規模的簡單工作。以下是 Gemini 3.1 Flash-Lite 最適合的用途：

• 翻譯：快速、便宜且大量翻譯，例如大規模處理即時通訊訊息、評論和支援服務單。你可以使用系統指令，將輸出內容限制為僅有譯文，不含額外註解：

  text = "Hey, are you down to grab some pizza later? I'm starving!"
  
  response = client.models.generate_content(
      model="gemini-3.1-flash-lite-preview",
      config={
          "system_instruction": "Only output the translated text"
      },
      contents=f"Translate the following text to German: {text}"
  )
  
  print(response.text)
  

• 轉錄：處理錄音、語音記事或任何音訊內容，並取得文字轉錄稿，不必另外啟動語音轉文字管道。支援多模態輸入，因此您可以直接傳遞音訊檔案進行轉錄：

  # URL = "https://storage.googleapis.com/generativeai-downloads/data/State_of_the_Union_Address_30_January_1961.mp3"
  
  # Upload the audio file to the GenAI File API
  uploaded_file = client.files.upload(file='sample.mp3')
  
  prompt = 'Generate a transcript of the audio.'
  
  response = client.models.generate_content(
    model="gemini-3.1-flash-lite-preview",
    contents=[prompt, uploaded_file]
  )
  
  print(response.text)
  

• 輕量型代理程式工作和資料擷取：支援實體擷取、分類和輕量型資料處理管道，並以結構化 JSON 輸出。舉例來說，從電子商務顧客評論中擷取結構化資料：

  from pydantic import BaseModel, Field
  
  prompt = "Analyze the user review and determine the aspect, sentiment score, summary quote, and return risk"
  input_text = "The boots look amazing and the leather is high quality, but they run way too small. I'm sending them back."
  
  class ReviewAnalysis(BaseModel):
      aspect: str = Field(description="The feature mentioned (e.g., Price, Comfort, Style, Shipping)")
      summary_quote: str = Field(description="The specific phrase from the review about this aspect")
      sentiment_score: int = Field(description="1 to 5 (1=worst, 5=best)")
      is_return_risk: bool = Field(description="True if the user mentions returning the item")
  
  response = client.models.generate_content(
      model="gemini-3.1-flash-lite-preview",
      contents=[prompt, input_text],
      config={
          "response_mime_type": "application/json",
          "response_json_schema": ReviewAnalysis.model_json_schema(),
      },
  )
  
  print(response.text)
  

• 處理文件和摘要：剖析 PDF 並傳回簡潔摘要，例如建構文件處理管道或快速分類收到的檔案：

  import httpx
  
  # Download a sample PDF document
  doc_url = "https://storage.googleapis.com/generativeai-downloads/data/med_gemini.pdf"
  doc_data = httpx.get(doc_url).content
  
  prompt = "Summarize this document"
  response = client.models.generate_content(
      model="gemini-3.1-flash-lite-preview",
      contents=[
          types.Part.from_bytes(
              data=doc_data,
              mime_type='application/pdf',
          ),
          prompt
      ]
  )
  
  print(response.text)
  

• 模型路徑：使用低延遲和低成本模型做為分類器，根據工作複雜度將查詢轉送至適當的模型。這是實際的生產模式，開放原始碼的 Gemini CLI 會使用 Flash-Lite 分類工作複雜度，並據此將工作轉送至 Flash 或 Pro。

  FLASH_MODEL = 'flash'
  PRO_MODEL = 'pro'
  
  CLASSIFIER_SYSTEM_PROMPT = f"""
  You are a specialized Task Routing AI. Your sole function is to analyze the user's request and classify its complexity. Choose between `{FLASH_MODEL}` (SIMPLE) or `{PRO_MODEL}` (COMPLEX).
  1.  `{FLASH_MODEL}`: A fast, efficient model for simple, well-defined tasks.
  2.  `{PRO_MODEL}`: A powerful, advanced model for complex, open-ended, or multi-step tasks.
  
  A task is COMPLEX if it meets ONE OR MORE of the following criteria:
  1.  High Operational Complexity (Est. 4+ Steps/Tool Calls)
  2.  Strategic Planning and Conceptual Design
  3.  High Ambiguity or Large Scope
  4.  Deep Debugging and Root Cause Analysis
  
  A task is SIMPLE if it is highly specific, bounded, and has Low Operational Complexity (Est. 1-3 tool calls).
  """
  
  user_input = "I'm getting an error 'Cannot read property 'map' of undefined' when I click the save button. Can you fix it?"
  
  response_schema = {
    "type": "object",
    "properties": {
      "reasoning": {
        "type": "string",
        "description": "A brief, step-by-step explanation for the model choice, referencing the rubric."
      },
      "model_choice": {
        "type": "string",
        "enum": [FLASH_MODEL, PRO_MODEL]
      }
    },
    "required": ["reasoning", "model_choice"]
  }
  
  response = client.models.generate_content(
      model="gemini-3.1-flash-lite-preview",
      contents=user_input,
      config={
          "system_instruction": CLASSIFIER_SYSTEM_PROMPT,
          "response_mime_type": "application/json",
          "response_json_schema": response_schema
      },
  )
  
  print(response.text)
  

• 思考：如要提高準確度，請針對需要逐步推論的工作設定思考，讓模型在產生最終輸出內容前，花費額外的運算資源進行內部推論：

  response = client.models.generate_content(
      model="gemini-3.1-flash-lite-preview",
      contents="How does AI work?",
      config=types.GenerateContentConfig(
          thinking_config=types.ThinkingConfig(thinking_level="high")
      ),
  )
  
  print(response.text)