AI_Qwen_text

The text-only sibling of AI_Qwen, hitting the same Alibaba Cloud DashScope API without the image inputs. What sets this one apart from the pack's other text-chat nodes is the model list itself: alongside general-purpose qwen3.5-plus and qwen3.5-flash, llm_model also offers qwen3-coder-plus, qwen3-coder-flash, and qwen3-coder-30b-a3b-instruct - Qwen's code-specialized line. If your use case inside a ComfyUI graph is actually generating or fixing code (a Python expression node, a JSON snippet, a small script) rather than prose, this is the node in the pack built for that, not the general GLM/Ollama text nodes.

Required fields: llm_model, preset (defaults to "None" until you've saved one with AI_PresetSave), custom_system_prompt, and text, your prompt. Optional: max_tokens (10–8192, default 1024), seed, api_key_input, and custom_model for a free-text override outside the five listed options. Outputs: result and system_prompt, the latter useful for confirming exactly what system prompt a preset actually injected.

Qwen's coder models sit inside the same broader Alibaba Tongyi push that made Apache-licensed, uncensored open weights the community default across video (Wan), photorealism (Z-Image), and image editing (Qwen-Image) through 2025 - the coder line is less talked about in that story than the image/video side, but it's built on the same underlying investment and the same API surface.

Practically, that means the same node can do double duty in a ComfyUI graph: use a general model for prompt drafting and a coder model in the same node, on a different call, for something like generating a small JSON payload another node expects. You don't need a second node just because the task changed shape - swap llm_model and keep going.

Installing it: ComfyUI Manager, search "ComfyUI-Apt_Preset," or clone manually - cd ComfyUI/custom_nodes && git clone https://github.com/cardenluo/ComfyUI-Apt_Preset - then install.bat and restart. No local weights, pure API client.

Getting a key: enable DashScope on an Alibaba Cloud account and generate a key from the console. Keep it straight from any ModelScope token you're also using elsewhere in this pack - they're separate services with separate credentials, and mixing them up is the single most common reason one of these Alibaba-flavored nodes fails while the other works fine.

Troubleshooting: auth errors point at api_key_input first - confirm it's a DashScope key and current. If you're using one of the -flash variants and getting weaker results than expected on a genuinely hard coding task, that's the expected quality/speed tradeoff of the flash tier - step up to qwen3-coder-plus or the 30B instruct model for anything nontrivial. And if result comes back truncated mid-answer, raise max_tokens before assuming the model failed; 1024 is a reasonable default for a short reply but tight for a full function or a longer explanation.