Qwen

The qwen provider routes to Alibaba’s Qwen models via DashScope. Qwen exposes two compatible surfaces and crabllm uses both: an OpenAI-compatible endpoint for chat, streaming, and embeddings, and a native Anthropic-compatible endpoint for Anthropic-format passthrough. Authentication is a bearer token on both (the Anthropic endpoint also accepts x-api-key).

Configuration

[providers.qwen]       # section name doubles as kind; add `kind = "qwen"` only if you rename it
api_key = "${DASHSCOPE_API_KEY}"
models = ["qwen-max", "qwen-plus", "qwen-turbo"]

Defaults, no base_url needed (international / -intl region):

• OpenAI-compatible: https://dashscope-intl.aliyuncs.com/compatible-mode/v1
• Anthropic-compatible: https://dashscope-intl.aliyuncs.com/apps/anthropic/v1 (the /v1/messages path the Anthropic SDK / Claude Code resolve from base_url)

Set base_url to override the OpenAI-compatible origin — e.g. the mainland-China endpoint https://dashscope.aliyuncs.com/compatible-mode/v1.

Usage

Send requests in OpenAI format as usual — just set the model to a Qwen id:

curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen-max",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

Anthropic-format requests are served natively by Qwen’s Anthropic-compatible endpoint rather than translated.

Supported Endpoints

• Chat completions (streaming and non-streaming)
• Embeddings
• Anthropic Messages (native passthrough, streaming and non-streaming)
• Tool/function calling

Limitations

• The Anthropic-compatible endpoint serves only /v1/messages — it has no /v1/models discovery route, which some Anthropic-native tools expect.