Qwen Provider¶

Alibaba Cloud Model Studio serves the Qwen lineup behind an OpenAI-compatible Chat Completions API — qwen3.7-max, qwen3.7-plus, qwen3.6-flash, qwen3-coder-plus, qwen3-vl-plus — with 1M-token context windows on most of the line and image input on all but the flagship.

Because Model Studio speaks the OpenAI API verbatim, RoomKit's provider is a thin subclass of OpenAIAIProvider: message building, tool calling and streaming all come for free. Four things are Qwen's own, and they are what the subclass exists for.

Install¶

pip install roomkit[qwen-ai]

The extra is qwen-ai, not qwen — qwen-tts and qwen-asr are separate voice extras installing different packages.

Quick start¶

from roomkit import AIChannel, RoomKit
from roomkit.providers.qwen import QwenAIProvider, QwenConfig

provider = QwenAIProvider(
    QwenConfig(
        api_key="sk-...",        # from https://bailian.console.aliyun.com
        model="qwen3.7-max",
    )
)

kit = RoomKit()
ai = AIChannel("ai-assistant", provider=provider, system_prompt="You are helpful.")
kit.register_channel(ai)

See examples/qwen_ai.py for an interactive chat.

Choosing your endpoint¶

Model Studio runs several deployments and your key belongs to exactly one. The default is the international endpoint; everything else is a base_url override:

Deployment	`base_url`
International (default)	`https://dashscope-intl.aliyuncs.com/compatible-mode/v1`
China (Beijing)	`https://dashscope.aliyuncs.com/compatible-mode/v1`
US (Virginia)	`https://dashscope-us.aliyuncs.com/compatible-mode/v1`
Workspace-scoped, any region	`https://{WorkspaceId}.{region}.maas.aliyuncs.com/compatible-mode/v1`

The workspace form takes its id from the Model Studio console, which is why it cannot be a default — there is no correct value to guess.

Thinking¶

Qwen takes a boolean switch plus a token cap, both outside the OpenAI schema:

QwenConfig(
    api_key="sk-...",
    model="qwen3.7-plus",
    enable_thinking=True,   # None leaves the model's own default
)

Per turn, AIChannel(thinking_budget=2048) maps straight onto Qwen's thinking_budget — this is the one provider where RoomKit's budget is the vendor's parameter rather than an approximation of it. A budget of 0 switches thinking off. The trace comes back in a dedicated field and surfaces as AIResponse.thinking (and as StreamThinkingDelta while streaming).

reasoning_effort is not sent

The field is inherited from OpenAIConfig and deliberately ignored here. Model Studio accepts it only for the third-party DeepSeek models it also hosts — see the DeepSeek guide for talking to those directly.

Model discovery¶

Model Studio's OpenAI-compatible deployment serves /chat/completions and nothing else — there is no /v1/models. list_models() therefore returns RoomKit's offline catalog rather than the account's own set, which is also why that catalog carries every id a caller is expected to reach:

for m in await provider.list_models():
    print(m.id, m.context_window, m.supports_vision)

An id outside the catalog still works on the wire; it just reports context_window is None, and supports_vision defaults to True so an image is never silently dropped.

Scope: Qwen, not the gateway¶

Model Studio also hosts DeepSeek, Kimi, GLM and MiniMax on the same endpoint. This provider is scoped to the Qwen family it is named for — its catalog, prices and thinking switch are Qwen's. Reaching the others through it would work on the wire and report the wrong metadata for every one of them; use their own providers, or OpenRouterAIProvider if a single key across vendors is what you want.

Pricing¶

The catalog carries Alibaba's international list prices, which is what ModelPricing.price() costs a response against.

Promotions, regions and tiers

Alibaba runs near-permanent limited-time promotions, quotes different rates per deployment (Beijing bills qwen3.7-max roughly a third below international), and tiers several models by input length. ModelPricing carries one list rate and a single long-context threshold, so qwen3-coder-plus and qwen3-vl-plus — four and three bands respectively — report no price rather than one that would understate a long-context bill by up to 12x. Check the billing page when the figure has to be exact.