How to Integrate Large Models with LangChain: A Step‑by‑Step Tutorial

Core Modules and Architecture

LangChain defines six core modules: Model I/O (standardizes model inputs and outputs), Retrieval (loads, splits, and embeds external data), Chains (links modules to build applications), Memory (stores historical information), Agents (general AI agents), and Callbacks (hooks for logging, monitoring, and streaming). These modules are organized into a three‑layer architecture.

Environment Setup

Create and activate a Conda environment named langchainenv and install LangChain:

conda create -n langchainenv python=3.12
conda activate langchainenv
pip install langchain
pip show langchain

Large‑Model API Design in LangChain

LangChain supports major LLM providers (OpenAI, Qwen, Gemini, DeepSeek) and locally hosted models (Ollama, Vllm). The integration pattern is to install the provider‑specific dependency (if required) and call the unified function init_chat_model to obtain a model object, abstracting provider differences.

Connecting SiliconFlow Qwen3‑8B

Register on the SiliconFlow website and obtain an API key.

Use the OpenAI‑compatible request format; no extra package is required.

from langchain.chat_models import init_chat_model

model = init_chat_model(
    model="Qwen/Qwen3-8B",
    model_provider="openai",
    base_url="https://api.siliconflow.cn/v1/",
    api_key="YOUR_API_KEY"
)

question = "你好，请问你是"
result = model.invoke(question)
print(result)

The call returns an AIMessage object. LangChain defines three message types: SystemMessage , HumanMessage , and AIMessage , representing system prompts, user inputs, and model outputs respectively.

Connecting DeepSeek

Install the provider package: pip install langchain-deepseek Initialize the chat model:

from langchain.chat_models import init_chat_model

model = init_chat_model(
    model='deepseek-chat',
    model_provider='deepseek',
    api_key="YOUR_DEEPSEEK_API_KEY"
)

question = "你好，请介绍一下你自己"
result = model.invoke(question)
print(result)

Switch to the reasoning model by changing the model argument to 'deepseek-reasoner':

model = init_chat_model(
    model='deepseek-reasoner',
    model_provider='deepseek',
    api_key="YOUR_DEEPSEEK_API_KEY"
)

The returned AIMessage includes additional_kwargs (chain‑of‑thought) and response_metadata (token usage, etc.).

Other Model Integrations

For any supported model, obtain the model’s API key, install the corresponding third‑party package (e.g., Tongyi Qwen), then initialize with init_chat_model and call invoke. The full list of supported models and their packages is documented at https://python.langchain.com/docs/integrations/chat/.