Rfym21/Qwen2API

Qwen‑Proxy (Qwen2API)

What it is – A self‑hosted proxy that turns Alibaba’s Qwen chat services (https://chat.qwen.ai and the Qwen CLI) into an OpenAI‑compatible (and partially Anthropic‑compatible) HTTP API. It lets any client that talks to the OpenAI API – e.g. ChatGPT‑Next‑Web, LobeChat, LangChain, Claude‑Code, etc. – use Qwen models without code changes.

Key capabilities

• OpenAI API surface – /v1/models, /v1/chat/completions, /v1/images/*, /v1/videos, plus the Anthropic‑style /v1/messages endpoint.
• Function calling / tools – full support for OpenAI tools and Anthropic tool_use, including streaming incremental arguments and strict tool_choice=required handling.
• Multi‑account rotation – configure several Qwen accounts; the proxy will round‑robin them (each can have its own outbound proxy) to avoid IP‑based bans.
• Per‑account outbound proxy – HTTP, HTTPS, SOCKS5/SOCKS5H proxies can be attached to individual accounts.
• 256 K context CLI endpoint – the /cli route exposes Qwen CLI models with very large context windows and tool support.
• Multimodal – image/video generation, image editing, and video understanding are exposed through the OpenAI‑style endpoints.
• Web UI – a built‑in management dashboard for adding accounts, setting per‑account proxies, monitoring usage, and editing model mappings.
• Persistence options – in‑memory (default), JSON file, or Redis for storing account data, settings, and logs.
• Docker & binary distribution – official Docker image, Docker‑Compose files, and a single‑file Bun binary that bundles the runtime, backend, frontend, and tokeniser.

How to get started

1. Docker (quickest)

   docker run -d \
     -p 3000:3000 \
     -e API_KEY=sk-admin123,sk-user456 \
     -e ACCOUNTS=user1@mail.com:pass1|http://proxy:8080 \
     -e DATA_SAVE_MODE=none \
     -v $(pwd)/caches:/app/caches \
     --name qwen2api rfym21/qwen2api:latest
   

2. Docker‑Compose – pull the provided docker-compose.yml and run docker compose up -d.
3. Local source – clone the repo, install with Bun (bun install --frozen-lockfile), copy .env.example to .env, edit the variables, then bun run start.
4. Standalone binary – run bun run build:binary to produce a single executable (qwen2api‑windows‑x64.exe, qwen2api‑linux‑x64, etc.) that can be started directly after placing a .env file next to it.

Configuration highlights

Variable	Purpose
API_KEY	Comma‑separated list of API keys. The first key is an admin key (access to the UI); the rest are normal keys for API calls.
ACCOUNTS	List of Qwen accounts in email:password form, optionally followed by `
PROXY_URL	Global outbound proxy used when an account does not define its own.
DATA_SAVE_MODE	none (default), file, or redis – where account data and settings are persisted.
CACHE_MODE	default (in‑memory) or file – where image cache files are stored.
QWEN_CHAT_PROXY_URL / QWEN_CLI_PROXY_URL	Override the upstream Qwen endpoints if you run your own mirrors.
AGENT_TURN_MAX_ATTEMPTS	Max retries for a tool‑calling round before returning an error.
MODELS_CACHE_TTL	Seconds to cache the upstream model list (0 = never expire).

Typical usage (OpenAI‑compatible client)

curl -H "Authorization: Bearer sk-admin123" \
     -H "Content-Type: application/json" \
     -d '{"model":"Qwen3.6-Plus","messages":[{"role":"user","content":"你好"}],"stream":false}' \
     http://localhost:3000/v1/chat/completions

The response follows the OpenAI schema (id, object, choices, usage).

Why you might need it

• Access Qwen’s powerful multimodal and 256 K‑context models while reusing existing OpenAI‑based tooling.
• Share a single Qwen subscription among multiple users or services without exposing raw credentials.
• Bypass IP‑based rate limits by rotating accounts and per‑account proxies.
• Run locally or in a container, keeping all traffic inside a trusted network.

Limitations / notes

• The proxy does not provide the original Qwen SDK; it merely forwards requests and adapts responses.
• Some Anthropic fields are only partially supported – the bridge aims for best‑effort compatibility, not a full Anthropic implementation.
• Persistent storage (file or Redis) is optional; without it, account tokens are kept only in memory and lost on restart.
• The service assumes a trusted environment – proxy URLs are returned verbatim in API responses.

---

TL;DR – Qwen‑Proxy is a Bun‑based server that lets any OpenAI‑compatible client talk to Alibaba’s Qwen models, with support for function calling, multimodal generation, account rotation, and a web UI. Deploy via Docker, Docker‑Compose, or a single binary, configure your Qwen accounts and API keys, and start using the familiar OpenAI endpoints.