💯SiliconCloudReasoning

This is the sibling node in the same pack as DeepSeekOnline, but it's a different front door. Instead of talking to DeepSeek's own API, it calls SiliconFlow's hosted endpoint (api.siliconflow.cn/v1) - a Chinese model aggregator where one key unlocks a zoo: DeepSeek R1 and V3, the whole Qwen2.5 family, Llama 3, GLM, Marco-o1, even the R1 distill sizes down to 1.5B. If you want DeepSeek's visible reasoning chain without dealing with DeepSeek's direct billing, or you want to A/B a dozen models without touching a single config, this is the node you'd reach for.

The use case is the same as its sibling: LLM-assisted prompting as a node. You describe what you want, the model writes a structured image prompt, and you wire the result straight into a CLIP Text Encode. On LLM-encoded checkpoints the encoder reads your prompt as an instruction, so a model writing that instruction for you is one LLM talking to another - and it's a workflow the community has genuinely converged on.

How it works

OpenAI-compatible chat client, same skeleton as DeepSeekOnline. Two dropdowns do the model selection: reasoning_model for R1 and its distills, model for everything else. The code refuses to run until you pick at least one - it returns a friendly "Please select a reasoning model or a model" string in the content output rather than crashing. One real design oddity: system and prompt get concatenated into a single user message (system + "," + prompt), so the system prompt isn't a separate role the model can weigh differently - keep it short. Unlike the sibling, this node exposes the full sampling kit: temperature, top_p, frequency_penalty, presence_penalty, max_tokens, and stream. The prefix_continuation and fim fields are passed through as an extra_body parameter.

Inputs and outputs that matter

For a beginner, pick a reasoning model or a model (or both - a reasoning model is used if one is selected), type your request into prompt, and leave the penalties alone. The defaults are sane.

Outputs: content is the raw reply - wire that into a CLIP Text Encode (Positive). reasoning captures the chain-of-thought when you've selected an R1 model, which is fun to watch stream in and useful for debugging what the model actually decided. format and negative prompt are where it gets more useful than the DeepSeek node: the pack parses the reply for a **Prompt:** marker into format, and a **Negative prompt:** marker into negative prompt - so if you prompt the model to emit both, you get a genuine negative out of one call. The catch is the parsing is regex-based and brittle: if the model doesn't emit those exact markers, format falls back to the whole reply and negative prompt comes back empty.

Installing it

Same pack, same steps as DeepSeekOnline. ComfyUI Manager → search ComfyUI-DeepSeek_Online, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Chengym2023/ComfyUI-DeepSeek_Online

Restart ComfyUI. The only dependency is openai - no model files, no heavy deps. Then edit config.json, and note the same trap as its sibling: the README's flat {"api_key": ...} example is wrong, the code reads a nested block. Use the shipped shape:

{
  "SiliconCloud": {
    "api_key": "your-siliconflow-api-key",
    "url": "https://api.siliconflow.cn/v1"
  }
}

Your key comes from cloud.siliconflow.cn (the README's registration link is the same door). If the node throws "Please add your SiliconCloud API key", that's the assert firing because the block is missing or empty.

Gotchas

The model lists are hardcoded in the pack source, so a newly released model won't show up until the author updates the pack - the 37-entry model dropdown is already a snapshot of early 2025. The Pro/-prefixed entries are SiliconFlow's pricier hosted tier; the plain ones are the cheap option that's plenty for prompt expansion. And remember: no system-role separation, brittle marker parsing on the derived outputs, and if you haven't selected a model at all, the node returns an error string to your content output rather than a hard failure. All three are quirks you live with, not reasons to avoid it - for the price, this node is a remarkably cheap way to put a real LLM inside your workflow.