RunningHub DeepSeek Chat

If you've ever wanted a reasoning LLM to rewrite your prompt or caption an image inside the graph, but your card is already full of a diffusion model, this is the node that sidesteps the whole problem. RunningHub DeepSeek Chat is a thin HTTP client that calls DeepSeek's V4 models through RunningHub's OpenAI-compatible API. It doesn't download weights, it doesn't touch your VRAM, and it runs the same whether you're on a 4090 or a potato. All it needs is an API key.

It lives in ComfyUI-WBLESS, LaoMaoBoss's grab-bag pack that also ships switches, scaling math, global variables and a whole family of API chat nodes (Gemini, Qwen, DeepSeek proper, and a running-Hub image one). This is the DeepSeek-focused member of that family.

Why you'd reach for it

Two honest reasons. First, DeepSeek V4 (the flash/pro tiers this node's dropdown exposes) has no open weights you're realistically running locally, so calling it is the only door. Second - and this is the one that gets people - if you run ComfyUI on RunningHub's own cloud platform, you can't just fire up a local Ollama, because it's a shared environment and localhost:11434 isn't yours. People post that exact question on r/comfyui. An API relay node is the answer the platform expects you to use.

The KB's framing applies here: for prompt enhancement specifically, a local 8B model is usually the smarter call - it's uncensored and free per call. This node is for when you want frontier-chat quality, you're on the cloud, or you just don't want to babysit a second model's VRAM.

How it works

Boring in the best way. The node builds an OpenAI-style /chat/completions request with your system_instruction and user_input as messages, POSTs it to {base_url}/chat/completions with Authorization: Bearer <your key>, and parses the JSON reply. The whole implementation is Python's stdlib urllib - no openai SDK, no pip dependency to break your environment, and the source is short enough to read in a minute.

It's a real API call, which is the double-edged part of the LLM-node category the KB flags: an API node holds a credential and phones home by design, and this category already shipped malware once (LLMVISION). This one is open source and tiny - skim core/runninghub_deepseek_chat.py before first use if you're at all wary. Also note: the code deliberately disables SSL certificate verification to survive corporate proxies and self-signed chains. That's a real trade, so treat your key like anything you paste into a node.

The inputs and outputs that matter

Set these and you're done:

• api_key - from RunningHub's enterprise API page (runninghub.ai/enterprise-api/sharedApi). "Bearer" prefix is optional; the node adds it for you.
• model - deepseek/deepseek-v4-flash or deepseek/deepseek-v4-pro. Flash is cheap and snappy; Pro is the heavier thinker.
• reasoning_effort - none is the default and skips the chain-of-thought. Flip to low/high/max when you actually want the model to think, and it'll show up in the reasoning output.
• user_input - what you're asking. The obvious one.
• base_url - leave it unless you're in China, where the README says switch to https://llm.runninghub.cn/v1.

max_tokens, temperature, top_p, presence_penalty and frequency_penalty are all standard sampling knobs; the defaults are fine for prompt work.

The three outputs wire straight into the graph: text (the model's answer - feed it into a prompt or text node), reasoning (the scratch-work, when effort is above none), and raw_response (the untouched JSON, your debugging friend).

Installing it

Install the pack once, get this node free:

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

Or skip the terminal: open ComfyUI Manager, search ComfyUI-WBLESS, install, restart. No model downloads, no pip install - the pack bundles its only helper (cozy_comfyui) locally.

Common issues

The big one is a missing key, and it fails loudly: the node raises "please provide a valid RunningHub API Key" the moment it runs. Check the enterprise/shared-API key, not a personal account one. Second gotcha: when the API returns an error object, the node doesn't crash - it writes API error: ... into the text output and moves on, which can look like a successful generation. When something seems subtly wrong, read raw_response. And if you're getting odd results from the CN mirror, confirm your key matches whichever base URL you chose.