Unified Client

OpenAI, DeepSeek, and Qwen all speak the OpenAI wire format, which means one openai Python client covers all three — you only swap base_url and model. This page wraps that insight into a small helper so your application code doesn't have to care which provider is underneath.

The three providers

Provider	base_url	Env var	A good default model
OpenAI	(default) https://api.openai.com/v1	OPENAI_API_KEY	gpt-4o-mini
DeepSeek	https://api.deepseek.com	DEEPSEEK_API_KEY	deepseek-chat
Qwen	https://dashscope.aliyuncs.com/compatible-mode/v1	DASHSCOPE_API_KEY	qwen-plus

A minimal unified function

unified_chat.py

import os
from dataclasses import dataclass
from openai import OpenAI


@dataclass(frozen=True)
class ProviderConfig:
    base_url: str | None   # None → use the SDK's default (OpenAI)
    env_var: str
    default_model: str


PROVIDERS: dict[str, ProviderConfig] = {
    "openai": ProviderConfig(
        base_url=None,
        env_var="OPENAI_API_KEY",
        default_model="gpt-4o-mini",
    ),
    "deepseek": ProviderConfig(
        base_url="https://api.deepseek.com",
        env_var="DEEPSEEK_API_KEY",
        default_model="deepseek-chat",
    ),
    "qwen": ProviderConfig(
        base_url="https://dashscope.aliyuncs.com/compatible-mode/v1",
        env_var="DASHSCOPE_API_KEY",
        default_model="qwen-plus",
    ),
}


def chat(
    prompt: str,
    provider: str = "openai",
    model: str | None = None,
    system: str | None = None,
) -> str:
    cfg = PROVIDERS[provider]
    client = OpenAI(
        api_key=os.environ[cfg.env_var],
        base_url=cfg.base_url,
    )

    messages = []
    if system:
        messages.append({"role": "system", "content": system})
    messages.append({"role": "user", "content": prompt})

    resp = client.chat.completions.create(
        model=model or cfg.default_model,
        messages=messages,
    )
    return resp.choices[0].message.content

Usage:

print(chat("Say hi in one sentence."))                      # OpenAI
print(chat("Say hi in one sentence.", "deepseek"))          # DeepSeek
print(chat("Say hi in one sentence.", "qwen", "qwen-max"))  # Qwen

The only line that differs between calls is the provider argument — the actual request code is reused.

Pitfalls to watch for

• Feature parity is not guaranteed. Tool use, JSON mode, and vision input sometimes differ or are missing on DeepSeek / Qwen. Treat "OpenAI-compatible" as "the plain chat endpoint works" — verify anything beyond that per-provider.
• Rate limits are per provider. A unified client does not give you a unified quota. If one provider throttles you, retry logic needs to know which one failed.
• Model IDs drift. Provider catalogs change faster than this tutorial. Pass an explicit model when you need a specific one; fall back to cfg.default_model for convenience.

Next

• Streaming — return tokens as they're generated (the unified pattern extends naturally).
• Tool Use — let the model call your Python functions.