Baidu ERNIE API Complete Guide: Python Integration and Enterprise Deployment

What This Tutorial Covers

You will master the complete usage of the Baidu ERNIE API:

• Qianfan platform registration and configuration
• ERNIE 4.5 / ERNIE Speed / ERNIE Lite model selection
• Three Python SDK calling methods
• Function Calling + search augmentation
• Plugin system and enterprise deployment
• Cost optimization strategies

🎯 Baidu ERNIE has the highest market share in China with the strongest Chinese language understanding (leading in classical texts, idioms, finance, and legal domains). The Qianfan platform provides the most comprehensive enterprise AI application infrastructure.

---

Step 1: Meet the Baidu AI Model Family

Core Models

Model	Type	Context	Characteristics	Price (per 1K tokens)
ERNIE 4.5	Flagship	128K	Strongest Chinese comprehension, knowledge-enhanced	¥0.12
ERNIE 4.0 Turbo	Flagship	128K	High-performance reasoning	¥0.08
ERNIE Speed	Fast	8K	Low latency, high concurrency	¥0.004
ERNIE Lite	Lightweight	8K	Generous free tier	Free
ERNIE Tiny	Ultra-light	4K	Extremely low cost	¥0.001
ERNIE Character	Role-play	8K	Emotional conversation	¥0.012

💡 The Qianfan platform also provides text-to-image (ERNIE-ViLG), speech synthesis, document parsing, and more.

---

Step 2: Get Your API Key

1. Visit the Qianfan Platform and sign up
2. Complete real-name verification (Chinese phone number / enterprise verification)
3. Create an application → get your API Key + Secret Key
4. Set environment variables:

export QIANFAN_AK="Your-API-Key"
export QIANFAN_SK="Your-Secret-Key"

New users receive substantial free credits. ERNIE Lite is completely free.

---

Step 3: Install SDK and Basic Usage

pip install qianfan

Basic Chat

import qianfan
import os

# Use IAM authentication
chat = qianfan.ChatCompletion(
    ak=os.getenv("QIANFAN_AK"),
    sk=os.getenv("QIANFAN_SK"),
)

response = chat.do(
    model="ernie-4.5-8k",  # Or "ernie-speed-8k", "ernie-lite-8k"
    messages=[
        {"role": "system", "content": "You are a senior Python developer"},
        {"role": "user", "content": "Implement a thread-safe LRU cache in Python"}
    ],
    temperature=0.7,
    top_p=0.8,
)

print(response["result"])

Streaming Output

response = chat.do(
    model="ernie-speed-8k",
    messages=[{"role": "user", "content": "Describe three key stages in the development of China's AI industry"}],
    stream=True,
)

for chunk in response:
    print(chunk["result"], end="", flush=True)

---

Step 4: Function Calling (Tool Invocation)

import json

# Define tools
tools = [{
    "type": "function",
    "function": {
        "name": "get_stock_price",
        "description": "Query real-time A-share stock prices",
        "parameters": {
            "type": "object",
            "properties": {
                "code": {
                    "type": "string",
                    "description": "Stock ticker, e.g., 600519 (Kweichow Moutai)"
                }
            },
            "required": ["code"]
        }
    }
}]

def get_stock_price(code: str) -> dict:
    """Simulated stock query (use East Money / Sina API in production)"""
    stocks = {
        "600519": {"name": "Kweichow Moutai", "price": 1688.50, "change": "+2.3%"},
        "000858": {"name": "Wuliangye", "price": 156.80, "change": "-0.5%"},
        "300750": {"name": "CATL", "price": 228.00, "change": "+5.1%"},
    }
    return stocks.get(code, {"error": "Stock not found"})

def chat_with_tools(user_query: str) -> str:
    response = chat.do(
        model="ernie-4.5-8k",
        messages=[{"role": "user", "content": user_query}],
        tools=tools,
    )

    msg = response.body["choices"][0]["message"]

    if "function_call" not in msg:
        return msg["content"]

    func_call = msg["function_call"]
    func_name = func_call["name"]
    func_args = json.loads(func_call["arguments"])

    print(f"🔧 Calling: {func_name}({func_args})")

    if func_name == "get_stock_price":
        result = get_stock_price(**func_args)

    # Return the result to the AI
    second_response = chat.do(
        model="ernie-4.5-8k",
        messages=[
            {"role": "user", "content": user_query},
            {"role": "assistant", "content": None, "function_call": func_call},
            {"role": "function", "name": func_name, "content": json.dumps(result, ensure_ascii=False)},
        ],
    )

    return second_response["result"]

print(chat_with_tools("What's the current price of Kweichow Moutai? Has it gone up?"))

---

Step 5: Search Augmentation (Real-Time Info)

The Qianfan platform has built-in web search capability:

response = chat.do(
    model="ernie-4.5-8k",
    messages=[{"role": "user", "content": "What are today's latest tech news?"}],
    system="You are a news summary assistant. Answer based on search results.",
    enable_citation=True,  # 🔑 Enable search augmentation
    search_info={"search_mode": "auto"},
)

print(response["result"])
# Search results return citation sources
if "search_info" in response:
    for ref in response["search_info"].get("search_results", []):
        print(f"📎 {ref['title']}: {ref['url']}")

---

Step 6: OpenAI Compatibility Mode

Qianfan also supports OpenAI SDK compatibility:

pip install openai

from openai import OpenAI

client = OpenAI(
    api_key=os.getenv("QIANFAN_AK"),
    base_url="https://qianfan.baidubce.com/v2",
)

response = client.chat.completions.create(
    model="ernie-speed-8k",
    messages=[{"role": "user", "content": "Describe autumn in Beijing using a classical Chinese poem"}],
    temperature=0.9,
)

print(response.choices[0].message.content)

---

Step 7: Text-to-Image

from qianfan.resources import Image2Text

# Text-to-image
i2t = Image2Text()
response = i2t.do(
    prompt="A cute orange tabby cat sitting beneath the red walls of the Forbidden City, sunlight filtering through ginkgo leaves, oil painting style",
    model="stable-diffusion-xl",
    size="1024x1024",
    n=1,
)

print(response["data"][0]["url"])

---

Step 8: Cost Optimization

Model Selection Strategy

Scenario	Recommended Model	Cost
Dev/Test	ERNIE Lite	Free
High-concurrency API	ERNIE Speed	¥0.004/1K tokens
Complex analysis	ERNIE 4.5	¥0.12/1K tokens
Lightweight tasks	ERNIE Tiny	¥0.001/1K tokens

Optimization Tips

# 1. Use System Prompt caching (supported by Qianfan)
# 2. Limit max_tokens to avoid waste
response = chat.do(
    model="ernie-speed-8k",
    messages=[{"role": "user", "content": "Summarize this article"}],
    max_output_tokens=512,  # Limit output length
    temperature=0.3,        # Lower temperature for creative tasks
)

# 3. Use async for batch processing
import asyncio
from qianfan import AsyncChatCompletion

async def batch_process(prompts: list[str]) -> list[str]:
    async_chat = AsyncChatCompletion()
    tasks = [
        async_chat.ado(
            model="ernie-speed-8k",
            messages=[{"role": "user", "content": p}],
        )
        for p in prompts
    ]
    results = await asyncio.gather(*tasks)
    return [r["result"] for r in results]

prompts = ["Explain machine learning", "Explain deep learning", "Explain reinforcement learning"]
results = asyncio.run(batch_process(prompts))

---

Pricing Quick Reference

Model	Input (¥/1K tokens)	Output (¥/1K tokens)
ERNIE 4.5	0.12	0.12
ERNIE 4.0 Turbo	0.08	0.08
ERNIE Speed	0.004	0.004
ERNIE Lite	0	0
ERNIE Tiny	0.001	0.001
ERNIE Character	0.012	0.012

---

Comparison with Other Models

Dimension	ERNIE 4.5	DeepSeek V4	Qwen3
Chinese comprehension	⭐⭐⭐⭐⭐	⭐⭐⭐⭐	⭐⭐⭐⭐
Coding ability	⭐⭐⭐	⭐⭐⭐⭐⭐	⭐⭐⭐⭐
Knowledge breadth	⭐⭐⭐⭐⭐	⭐⭐⭐⭐	⭐⭐⭐⭐
API pricing	Moderate	Low	Moderate
Enterprise ecosystem	⭐⭐⭐⭐⭐	⭐⭐⭐	⭐⭐⭐⭐

ERNIE’s strengths are deep Chinese language understanding, enterprise ecosystem (Qianfan platform), and search augmentation. Its coding ability is not as strong as DeepSeek’s, but it’s irreplaceable for Chinese-language scenarios.

---

FAQ

Q: How do I choose between ERNIE and DeepSeek?

A: Choose ERNIE for Chinese knowledge-intensive tasks (legal/finance/classical texts); choose DeepSeek for coding/math/reasoning; choose Qwen for multimodal needs.

Q: Is the Qianfan platform’s free tier sufficient?

A: ERNIE Lite is completely free, and ERNIE Speed has generous free credits. Individual developers generally don’t need to pay.

---

Next Steps

• DeepSeek API Beginner Guide
• China AI Models Ultimate Comparison

📝 Based on Baidu Qianfan platform latest version tested in June 2026.