Qwen语言丨API

You want an LLM's judgment inside a ComfyUI graph - a prompt rewritten, a caption generated, a decision made - but you don't want to download a 30GB model or babysit VRAM. QwenLanguageAPI is the QING pack's hosted answer: it calls Qwen's language models through an OpenAI-compatible API and drops the text back into your workflow. No local GPU involved, just a key and a network connection. It's the LLM-as-node pattern from the KB's llm-in-comfyui doc, wired for Alibaba's Qwen family.

How it works

The node talks to Qwen through two providers, both OpenAI-compatible (the pack uses the openai client):

• 阿里云百炼 (DashScope) - base URL dashscope.aliyuncs.com/compatible-mode/v1, key env DASHSCOPE_API_KEY.
• 硅基流动 (SiliconFlow) - base URL api.siliconflow.cn/v1, key env SILICONFLOW_API_KEY.

Inputs: text_input (your prompt), platform, model (a dropdown with qwen3-max, qwen-plus, qwen-turbo, qwq-plus, and several sized variants), max_tokens (default 4096), and history (default 20 - how many conversation turns to keep in memory). Optionals: temperature (0.3), top_p (0.85), repetition_penalty (1.1), and clear_history (dump the cached conversation).

Outputs: generated_text, conversation_info (a summary of the dialogue state), and total_tokens (so you can watch your spend).

Two behaviors worth knowing: the conversation is cached per platform+model, so successive runs actually hold a multi-turn chat - until you set clear_history. And the model names are mapped to real API IDs under the hood (qwq-plus → qwq-32b-preview, for example), plus qwq reasoning models automatically get their temperature capped at 0.5 because they hate heat.

Where you'd reach for it

Prompt enhancement and rewriting are the flagship uses: rough idea in, structured model-appropriate prompt out, fed straight to your CLIP Text Encode. Captioning rough drafts, summarization, and any "ask a model a question and branch on the answer" logic all work - pair the output with QING_ModelFallback so a rate-limit failure hands the graph a backup model instead of a dead batch.

Installing and configuring

cd ComfyUI/custom_nodes
git clone https://github.com/GAO-SHIQING/ComfyUI-QING
cd ComfyUI-QING
python install_dependencies.py

Restart ComfyUI after. The install script pulls the openai client, which this node needs. Then the key: Settings → QING → API configuration (the pack's config panel), or set the env vars above. Keys are stored locally and auto-injected at runtime - they never sit in your workflow JSON. Grab a DashScope key at bailian.console.aliyun.com or a SiliconFlow key at cloud.siliconflow.cn. (README's GAOSHI-QING clone-URL typo - the repo is GAO-SHIQING/ComfyUI-QING.)

Things to know

It's a network call - needs internet, needs the key, and SiliconFlow caps max_tokens at 4096 while DashScope supports more, so the same workflow behaves differently per platform. Watch total_tokens if you're worried about cost; a long history multiplies the tokens on every call. And the usual LLM-in-a-pipeline caveat applies: the output is freeform text, so if you're parsing it programmatically (JSON out, decision values), a chat model's rambling will bite you - scope the request narrowly and validate the result downstream.