Deepseek Chat

Let's get the big thing out of the way: this node does not run DeepSeek on your machine. There's no model download, no GPU requirement, no VRAM math. "Deepseek Chat" is a thin wrapper that takes your prompt, phones DeepSeek's API (the deepseek-chat model, i.e. V3), and hands the reply back to your graph as a plain text string. That's the entire job - and for a lot of workflows, it's exactly what you want.

Why bother putting an LLM in ComfyUI at all? Because ComfyUI is a graph, and graphs eat text. The classic move: use the node to generate or refine a prompt from a short description, then feed the output string straight into a CLIPTextEncode. It's also handy for translating, brainstorming variations, or acting as a little assistant while you build a workflow. Anything where you want a model's words to become an input to your pipeline.

How it works

On execution, the node reads your API key from a config.json inside the pack folder, spins up an OpenAI client pointed at api.deepseek.com, and sends a chat completion with your prompt plus a system_prompt. The reply comes back as the node's single STRING output. No conversation memory - every run is a fresh, single-shot call.

The inputs that matter

There are only three, and that's the point of this node - it's the "just talk to the model" option in the pack.

• prompt (multiline) - the actual question. The only thing you'll edit every run.
• system_prompt (multiline, default "You are a helpful assistant") - the persona/instructions. Set it once, forget it.
• temperature (0–2, default 0.7) - the one knob. Lower = stricter, higher = more creative.

One output, STRING, which you can wire into a text encoder, a display, or anything that takes text.

Install

The pack is small and light - dependencies are just openai and requests, and there are no models to download.

cd ComfyUI/custom_nodes
git clone https://github.com/yichengup/Comfyui-Deepseek

Or search "Comfyui-Deepseek" in ComfyUI Manager and hit install. Then restart ComfyUI.

The part that trips everyone up: the API key lives in custom_nodes/Comfyui-Deepseek/config.json, not in any node field. Grab a key from platform.deepseek.com and edit the file:

{
  "api_key": "sk-your-real-key-here",
  "base_url": "https://api.deepseek.com",
  "silicon_api_key": "key值",
  "silicon_base_url": "https://api.siliconflow.cn/v1"
}

The key is read when the node instance is created, so edit the file, then restart ComfyUI (or at least reload the workflow). Leaving the placeholder key值 in place won't work.

Common issues

• Output says Error: Please configure your API key in config.json - key is missing or still the placeholder. Fix config.json and restart.
• Any other Error: ... string - that's the API complaining (wrong key, empty balance, rate limit) surfacing as the node's output. Remember: with this pack, errors come back as text, not as a thrown error, so don't mistake one for a bad answer.
• No memory - this node forgets everything between runs. If you want a model that remembers the conversation, the pack's "Deepseek Reasoner" node keeps history.

One honest note: this is a small, barely-documented pack (the README is basically three lines and two links), so don't expect polish. It's a functional API postcard, nothing more.