langchain-qwq-modification

魔改langchain-qwq，提高兼容性：

1. 为api_base设置别名base_url；
2. 合并reasoning_content到content；
3. 添加与标签

Installation

pip install -U langchain-qwq-modification

---

以下为原版langchain-qwq的readme

---

langchain-qwq

This package contains the LangChain integration with QwQ and Qwen3

Installation

pip install -U langchain-qwq

OR if you want to install additional dependencies when you clone the repo:

pip install -U langchain-qwq[docs]
pip install -U langchain-qwq[test]
pip install -U langchain-qwq[codespell]
pip install -U langchain-qwq[lint]
pip install -U langchain-qwq[typing]

Environment Variables

And you should configure credentials by setting the following environment variables:

• DASHSCOPE_API_KEY: Your DashScope API key for accessing QwQ or Qwen3 models
• DASHSCOPE_API_BASE: (Optional) API base URL, defaults to "https://dashscope-intl.aliyuncs.com/compatible-mode/v1"

Chat Models (QwQ)

ChatQwQ class exposes chat models from QwQ. The integration works directly with a standard API key without requiring the Tongyi dependency.

from langchain_qwq import ChatQwQ

llm = ChatQwQ()
llm.invoke("Sing a ballad of LangChain.")

Advanced Usage

Streaming

llm = ChatQwQ(model="qwq-plus")
for chunk in llm.stream("Write a short poem about AI"):
    print(chunk.content, end="")

Async Support

llm = ChatQwQ(model="qwq-plus")
response = await llm.ainvoke("What is the capital of France?")
print(response.content)

# Streaming
async for chunk in llm.astream("Tell me about quantum computing"):
    print(chunk.content, end="")

Access to Reasoning Content

response = llm.invoke("Explain how photosynthesis works")
content = response.content
reasoning = response.additional_kwargs.get("reasoning_content", "")

Merge Reasoning Content to Content

from langchain_qwq import ChatQwQ
from langchain_qwq.utils import convert_reasoning_to_content
model = ChatQwQ(
    model="qwq-plus"
)
for chunk in convert_reasoning_to_content(model.stream("hello")):
    print(chunk)

also support async

from langchain_qwq.utils import aconvert_reasoning_to_content
async for chunk in aconvert_reasoning_to_content(model.astream("hello")):
    print(chunk)

and you can custom the think tag.

from langchain_qwq.utils import convert_reasoning_to_content

async for chunk in aconvert_reasoning_to_content(
        model.astream("hello"), think_tag=("<Start>", "<End>")
    ):
        print(chunk)

Tool Calls

from langchain_core.tools import tool

@tool
def get_current_weather(location: str, unit: str = "fahrenheit"):
    """Get the current weather in a given location"""
    return f"72 degrees and sunny in {location}"

llm = ChatQwQ(model="qwq-plus")
llm_with_tools = llm.bind_tools([get_current_weather])
response = llm_with_tools.invoke("What's the weather in San Francisco?")

also you can use parallel_tool_calls to enable parallel tool calls

from langchain_qwq import ChatQwQ
from langchain_core.tools import tool


@tool
async def get_data_in_db(db_name: str) -> str:
    """get the data from the database"""
    return f"data from the database {db_name}"


model = ChatQwQ(model="qwq-plus").bind_tools([get_data_in_db])

print(
    model.invoke(
        "please get the data from the database user and animal",
        extra_body={"parallel_tool_calls": True},
    )
)

Chat Models (Qwen3)

You can call Qwen3 through ChatQwQ, but we more strongly recommend using ChatQwen to call Qwen3 as it provides better support.

there is some examples code

from langchain_qwq import ChatQwen


model = ChatQwen(model="qwen3-32b")

for chunk in model.stream("hello!"):
    print(chunk)

Qwen3 is hybrid reasoning, so you can use enable_thinking to turn off the thinking process.

from langchain_qwq import ChatQwen
from dotenv import load_dotenv


model = ChatQwen(model="qwen3-32b", enable_thinking=False)

for chunk in model.stream("hello!"):
    print(chunk)

Qwen3 also supports thinking_budget to control the length of the thinking process.

from langchain_qwq import ChatQwen
from dotenv import load_dotenv

model = ChatQwen(model="qwen3-32b", thinking_budget=100)

for chunk in model.stream("hello!"):
    print(chunk)

Qwen3's integration is through ChatQwen, which inherits from QwQ, so you can use any features supported by ChatQwQ such as async, etc.