⛱️Qwen2 Chat

⛱️Qwen2 Chat is the multi-turn sibling of ⛱️Qwen2 in the same pack: identical inputs, but instead of forgetting every exchange, it keeps a running transcript and returns it. The name sells it as the "chat about your ideas" node, and that's the honest use - a back-and-forth where you refine an idea over a few runs rather than firing one shot and moving on.

But here's the thing to know before you use it: the output is the entire conversation, not the latest answer. The node appends each exchange to an in-memory chat_history on the node instance and its text output is the whole thing, formatted as role: content lines with a ----- separator between turns. If you plug that into a CLIPTextEncode prompt box expecting a fresh prompt, you'll get three turns of rambling conversation instead. For actual prompt generation, the single-turn ⛱️Qwen2 node is the one you want; use this one when you genuinely want the transcript, or want to carry an idea through several iterations before you copy the result somewhere.

Reality check on "context multi-turn"

The README bills this as "上下文多轮对话" - context multi-turn chat - and the source tells a slightly different story. Each run builds the message list fresh from just the system instruction plus your current prompt; the chat history is never fed back into the model. So it's a running transcript, not real memory. The model doesn't know what you said two turns ago; it just shows you everything it has ever said in one string. That's fine for a visible conversation log, but don't rely on it for coherent multi-step reasoning across turns.

Two more quirks, both straight from the code. The history lives on the node instance, so it accumulates across queue runs with no clear button - if it starts answering like it's stuck in an old conversation, replace the node or restart ComfyUI. And responses get logged under the role system rather than assistant, so the transcript labels look a little off. Harmless, but it'll confuse you once if you're not expecting it.

Inputs and outputs

Same four required inputs as its single-turn sibling:

• model and tokenizer - from the ⛱️Qwen2 ModelLoader's outputs; both must be connected.
• prompt - your current question or request, multiline.
• system_instruction - the persona. The default is the same prompt-embellishment instruction, so this node works fine for iterative prompt drafting too.

One output, text: the formatted transcript as a plain STRING, ready for a display node like DisplayText_Zho (from the same author's ComfyUI-Gemini pack) or any text input that wants it.

Install

Same pack, same steps - ComfyUI Manager, search ComfyUI-Qwen, or:

cd ComfyUI/custom_nodes
git clone https://github.com/ZHO-ZHO-ZHO/ComfyUI-Qwen
cd ComfyUI-Qwen
pip install -r requirements.txt

Restart after installing. The only dependency is transformers>=4.40.0, and the model is auto-downloaded on first run by the ModelLoader. Stick to the 0.5B–7B Qwen2.5 Instruct sizes on a normal GPU; generation happens entirely locally, with no API and no key.

Bottom line

It's a nice little transcript-keeping chat node, and the underlying Qwen models are solid for this job. Just remember the output is the conversation, not the answer - and that "context" is in the transcript only. If you wanted a prompt generator, grab the single-turn node; if you wanted a local chatbot with real memory, this isn't that, either.