Qwen-Reverse

  

Reverse-engineered async Python client for chat.qwen.ai — text chat, streaming with real-time reasoning, tool calling, image and video generation. No official API key required (works anonymously).

pip install qwen-reverse

---

⚡ Performance, Limits & Benchmarks (Tested & Verified)

In-depth stress testing on the live chat.qwen.ai backend reveals the following real-world limits and throughput capabilities:

📊 Concurrency & Rate Limit Results

Concurrent Requests	Success Rate	Avg Total Batch Time	Status	Notes
3 requests	3 / 3 (100%)	~4.98s	🟢 Passed	Flawless
5 requests	5 / 5 (100%)	~5.03s	🟢 Passed	Flawless
10 requests	10 / 10 (100%)	~4.11s	🟢 Passed	Flawless
20 requests	20 / 20 (100%)	~7.68s	🟢 Passed	Flawless
30 requests	30 / 30 (100%)	~10.19s	🟢 Passed	Flawless
50 requests	50 / 50 (100%)	~5.36s	🟢 Passed	Maximum safe burst per IP
80 requests	0 / 80 (0%)	1.35s	🔴 WAF Triggered	QwenError: WAF blocked chat creation

🛡️ Key Performance Takeaways

• Single IP Burst Limit: Up to 50 concurrent requests in parallel succeed with 100% reliability on a single IP without authentication.
• WAF Protection Trigger: Sending a sudden burst of ≥80 concurrent requests <1 second triggers Alibaba Cloud WAF IP throttling (QwenError: WAF blocked chat creation).
• WAF Cooldown Duration: An IP block typically cools down automatically in 3 to 5 minutes.
• Instant WAF Bypass via Proxy Rotation: Passing a proxy (proxy="http://ip:port") instantly bypasses any IP-level WAF cooldown block with 100% success rate.
• Infinite Scaling Strategy: By combining Proxy Rotation (proxy=...) with Account Token Rotation (SharedTokenManager), you can achieve virtually unlimited requests per minute (1,000+ RPM).

---

Features

• Chat — one-shot, streaming (SSE incremental), multi-turn with conversation memory (conversation_id / parent_id chaining)
• Real-time reasoning — reasoning events streamed token-by-token before the answer, in both plain and multi-turn mode
• Tool calling — OpenAI-style function definitions; either let the SDK execute them (JSON-stringified follow-up) or handle them yourself with emit_tool_calls=True
• Vision & Document Analysis — image and file upload (files=["photo.jpg", "doc.txt"]) for vision-capable models
• Image editing (i2i) — chat_type="image_edit": edit an uploaded image (add/remove/modify) and get back CDN URLs
• File upload — upload() uploads local files (or URLs/bytes) into the web API for vision/editing chats
• Image / video generation — t2i and t2v returning CDN URLs (cdn.qwenlm.ai)
• No account required — the web API works without a token; OAuth device-flow login (chat.qwen.ai account) is also implemented
• Anti-Bot & WAF Evasion — built-in BXUAGenerator (cryptographic bx-ua header generation), browser fingerprinting, and session cookie handling (ssxmod_itna)
• Optional FastAPI server — OpenAI-compatible /v1/chat/completions and /v1/models (see server/)

---

Quickstart

Basic Text Generation

import asyncio
from qwen_reverse import Generative

async def main():
    gen = Generative()
    print(await gen.generate("Explain what a kernel is in 2 lines"))

asyncio.run(main())

Streaming with Real-Time Thinking (Chain of Thought)

import asyncio
from qwen_reverse import Generative

async def main():
    gen = Generative()
    async for event in gen.stream("Explain in 2 sentences what an OS kernel is"):
        if event["type"] == "reasoning":
            print(f"\r\033[90m{event['data']}\033[0m", end="", flush=True)
        elif event["type"] == "content":
            print(event["data"], end="", flush=True)

asyncio.run(main())

Multi-turn Conversation Memory

import asyncio
from qwen_reverse import Conversation

async def main():
    conv = Conversation()
    r1 = await conv.send("My name is Popbob and I work with kernels in C.")
    r2 = await conv.send("What is my name?")  # remembers turn 1
    print(r2)  # "Popbob"
    print(conv.conversation_id, conv.parent_id)

asyncio.run(main())

Tool Calling (Automatic Execution)

import asyncio
from qwen_reverse import Conversation

WEATHER_TOOL = {
    "type": "function",
    "function": {
        "name": "get_weather",
        "description": "Get current weather for a city",
        "parameters": {
            "type": "object",
            "properties": {"city": {"type": "string"}},
            "required": ["city"],
        },
    },
}

async def main():
    conv = Conversation(tools=[WEATHER_TOOL])
    async for event in conv.stream("What's the weather in Buenos Aires?"):
        if event["type"] == "tool_calls":
            print("[tool_calls]", event["data"])
        elif event["type"] == "content":
            print(event["data"], end="", flush=True)

asyncio.run(main())

High Concurrency & Proxy Rotation Example

import asyncio
import itertools
from qwen_reverse import Conversation

# List of HTTP/SOCKS proxies
PROXIES = [
    "http://1.231.81.166:3128",
    "http://108.181.123.113:3128",
    "http://123.138.24.113:9443"
]
proxy_pool = itertools.cycle(PROXIES)

async def worker(req_id: int):
    proxy = next(proxy_pool)
    conv = Conversation(model="qwen3.7-plus", proxy=proxy, timeout=15)
    reply = await conv.send(f"Say hello to worker {req_id}")
    print(f"Worker {req_id} via {proxy}: {reply.strip()}")

async def main():
    # Execute 30 concurrent requests across rotated proxies
    tasks = [worker(i) for i in range(30)]
    await asyncio.gather(*tasks)

asyncio.run(main())

Image Generation

import asyncio
from qwen_reverse import Image

async def main():
    img = Image()
    urls = await img.generate(
        "A cyberpunk dragon flying over a neon city, anime style",
        aspect_ratio="16:9",
    )
    print(urls[0])  # https://cdn.qwenlm.ai/output/...

asyncio.run(main())

Upload a File + Chat with Vision

import asyncio
from qwen_reverse import Conversation

async def main():
    conv = Conversation(model="qwen3-vl-plus")  # a vision-capable model
    reply = await conv.send(
        "What is written on the whiteboard?",
        files=["photo.jpg"],  # path | bytes | URL | already-uploaded dict
    )
    print(reply)

asyncio.run(main())

---

API Reference

Symbol	Description
Generative(model=..., token=...)	.generate(), .stream() (events: reasoning, content, usage, tool_calls, done)
Conversation(token=..., tools=..., proxy=...)	.send(), .stream(), .reset() — persists conversation_id / parent_id across turns
Image(model=...), Video(model=...), ImageEdit(model=...)	.generate(prompt, aspect_ratio=...) → list of CDN URLs; ImageEdit.edit(prompt, image, aspect_ratio=...) for image-to-image editing
create_chat(model, messages, ...)	low-level async generator; conversation_id / parent_id for multi-turn; reasoning_effort accepts none / low / medium / high
fetch_models()	fetch available models (qwen3.8-max, qwen3.7-plus, qwen3-vl-plus, ...)
upload(data, filename=...)	upload a local path/bytes (or fetch+upload a URL) → file payload dict
resolve_files(files)	normalize a list of paths/bytes/URLs/dicts into upload-ready payload dicts
start_device_login() / complete_device_login(...)	OAuth device flow for authenticated accounts
SharedTokenManager	thread-safe token manager & token rotation across multiple accounts
BXUAGenerator	WAF anti-bot primitive generating bx-ua signatures

---

Authentication

Anonymous mode works — no token required for most features. For higher limits or video generation, log in with a chat.qwen.ai account:

import asyncio
from qwen_reverse import start_device_login, complete_device_login

async def main():
    client, data = await start_device_login()
    print(data["verification_uri_complete"])  # open in a logged-in browser
    token = await complete_device_login(client, data)
    print("Logged in token:", token)

asyncio.run(main())

---

Running Tests

pip install -e ".[dev]"
pytest

---

Optional OpenAI + Anthropic-Compatible Server

A FastAPI server that exposes chat.qwen.ai through standard APIs, so existing tools (Claude Code, OpenAI SDKs, Cursor, Gemini CLI, etc.) can use Qwen models without changes.

pip install -e .            # server deps (fastapi, uvicorn) are included
qwen-reverse                # starts server, picks a free port, asks which agent to launch
qwen-reverse --claude       # starts server + launches Claude Code pointed at it
qwen-reverse --model qwen3.8-max-thinking   # model override for the launched agent
python run.py               # same as `qwen-reverse` (repo checkout)

The CLI (qwen-reverse) detects installed agents on your PATH and can launch them with the right env vars: --claude, --gemini, --cursor, --qwen, --opencode, --aider, plus --openai to print the OpenAI-compatible setup, --port, --host, --no-server.

Endpoints

Endpoint	Protocol	Notes
POST /v1/chat/completions	OpenAI	streaming SSE + non-streaming, reasoning_content, tool calls
POST /v1/completions	OpenAI (legacy)	wraps chat completions with a plain prompt
POST /v1/embeddings	OpenAI	placeholder vectors (no real embedding backend)
GET /v1/models	OpenAI	base models + -search / -thinking / -web-dev / -deep-research / -artifacts / -slides variants
POST /v1/messages	Anthropic	streaming + non-streaming, translated to Anthropic SSE (thinking blocks)
POST /v1/images/generations	OpenAI	text-to-image → CDN URLs
POST /v1/images/edits	OpenAI	image editing (multipart or JSON) → CDN URLs
GET /health	—	liveness probe

Anthropic model names are mapped automatically (claude-opus-4-6 → qwen3.8-max), and model suffixes select special modes: -thinking (reasoning_effort=high), -search, -web-dev, -deep-research, -artifacts, -slides, -image/-t2i, -video/-t2v.

# manual start (repo checkout):
uvicorn qwen_reverse.server.app:app --host 127.0.0.1 --port 8090

Using it with Claude Code

export ANTHROPIC_BASE_URL=http://127.0.0.1:8090
export ANTHROPIC_API_KEY=qwen-reverse
export ANTHROPIC_MODEL=claude-opus-4-6-thinking   # or any claude-* name, or qwen names with suffixes
claude

---

How It Works

The client replicates what the official web app does against chat.qwen.ai:

1. Create a chat via POST /api/v2/chats/new (gets a chat_id)
2. Stream the response via POST /api/v2/chat/completions?chat_id=... with version: 2.1, incremental_output and feature_config enabling reasoning streaming
3. Multi-turn chaining uses the server's response_id (assistant message fid) as the next turn's parent_id
4. Anti-bot headers are regenerated per request: ssxmod_itna cookies (custom LZW + custom base64), bx-umidtoken, bx-ua signatures, and browser fingerprinting

---

Disclaimer

This project is for educational and research purposes. It is not affiliated with or endorsed by Alibaba/Qwen. Use at your own risk — the endpoints may change or the service may rate-limit or block unofficial clients. MIT licensed; reverse-engineering references based on g4f (MIT).