Deepseek Run

The name is a bit misleading - this doesn't call DeepSeek's cloud API, and there's no key to configure. The single model in its dropdown is DeepScaleR-1.5B-Preview, a small open reasoning model distilled and reinforcement-learning-tuned from DeepSeek-R1, and it runs entirely on your own GPU. Despite being tiny by LLM standards (1.5B parameters), models in this family are notable for punching well above their weight on step-by-step reasoning tasks - which matters for this node, because it's built specifically around a reasoning model's behavior rather than a plain instruct model's.

How it works

Reasoning models don't just answer - they "think" first, generating a chain of intermediate reasoning before committing to a final response. That's exactly why this node has two separate string outputs instead of one: Think carries the model's scratchpad reasoning, and Prompt carries the clean final answer. You don't have to manually strip a reasoning block out of the output before wiring it downstream - the node already splits it for you.

The inputs and outputs that matter

• model - currently a single choice, DeepScaleR-1.5B-Preview. The tooltip is specific and important: models are expected to already be sitting in your ComfyUI/models/LLM folder. This one is not pulled down automatically the way you might expect from the pack's Qwen nodes - you have to go get the weights yourself and put them there before this node has anything to select.
• user_prompt - your actual input to the model. Defaults to empty.
• temperature - note the range here is 0–2, wider than the 0–1 range on this pack's other LLM nodes. Push it up for more varied reasoning paths, keep it low for more literal, repeatable output.
• top_k and top_p - the usual sampling controls, same idea as elsewhere in this pack.
• max_tokens - response length cap.
• unload_model - off by default, meaning the model stays resident in VRAM between runs. The tooltip states plainly that turning it on unloads it after execution, and the next call pays a reload cost. Flip it on if you're tight on VRAM and only call this node occasionally.
• seed - reproducibility.

How to install it

ComfyUI Manager: search "ComfyUI_OneButtonPrompt", install, restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/billwuhao/ComfyUI_OneButtonPrompt_Flux

Restart ComfyUI. Then, separately, source the DeepScaleR-1.5B-Preview weights and place them in ComfyUI/models/LLM - the node's dropdown has nothing to offer until that folder exists and has the model in it. The README doesn't spell out a dependency file for this pack; if the node itself fails to import, check the console for a missing Python package and install it into ComfyUI's own environment.

Common issues

The model dropdown is empty, or the node errors immediately. This is almost always the missing-model-folder problem. Unlike this pack's other LLM nodes, this one won't fetch weights for you - go get DeepScaleR-1.5B-Preview and drop it into ComfyUI/models/LLM first.

Confusing this for a hosted API node. It isn't. There's no key field anywhere in the schema because there's no external call being made - everything runs locally against the model file you supplied.

Not sure which output to use. If you just want a usable prompt, wire Prompt, not Think - the Think output is the model's reasoning trail, useful for debugging or curiosity, not for feeding into a text encoder.