DeepSeek Chat

DeepSeek is the Chinese LLM lab whose cheap, open-ish models took the world by surprise, and this node is the pack's direct line to them. It's a straightforward chat node: system instruction plus user input in, text out, with the whole OpenAI-style response also available as a raw string. It's the cheapest of the pack's three chat nodes to run, and for prompt engineering, tag generation, or batch metadata it does the job without any local weights.

The standout feature is the reasoning output. The node ships with two model choices - deepseek-chat (the fast general model) and deepseek-reasoner (the reasoning model that thinks before it answers) - and when you pick the reasoner, its visible chain-of-thought comes back in a separate output. That's a genuinely useful hook: you can run a prompt through the reasoner, capture its reasoning, and use that to decide downstream behavior, or just inspect why it wrote what it wrote.

How it works

Under the hood it POSTs to DeepSeek's OpenAI-compatible endpoint (api.deepseek.com/chat/completions) with your messages and sampling settings. It's polite about the API key, accepting it with or without a Bearer prefix and adding it for you. The response is parsed into:

• text - the assistant's actual reply.
• reasoning - the reasoning_content from the reasoner model (empty when you're on deepseek-chat).
• raw_response - the full JSON, for debugging or extracting extra fields.

Inputs: api_key, model (the two-way enum), system_instruction (defaults to "You are a helpful assistant"), user_input, temperature (0–2), and top_p (0–1). No image input - this one is text-only, unlike the pack's Qwen and Gemini chat nodes.

Install

Part of ComfyUI-WBLESS:

cd ComfyUI/custom_nodes
git clone https://github.com/LaoMaoBoss/ComfyUI-WBLESS

restart ComfyUI, or search "ComfyUI-WBLESS" in ComfyUI Manager. No models, no dependencies - but you do need a paid DeepSeek API key from their platform.

Troubleshooting

A 401 means your key is wrong or the account isn't funded - DeepSeek is pay-as-you-go, and an empty balance fails with an auth-style error that sends people hunting the wrong problem. If reasoning is always empty, you're on deepseek-chat, which doesn't produce chain-of-thought; that's expected. Reasoner responses are slower and pricier per token than the plain chat model, so don't default to it for every call. And because everything runs on DeepSeek's servers, latency is their problem, not yours - but a long user_input with a reasoner model can feel very slow in the graph. For the price, this is the sensible default chat node in the pack for text-only work.