Silicon Deepseek Reasoner

If you want the whole DeepSeek-R1 package - the reasoning trace, the final answer, conversation history, and a system prompt - this is the kitchen sink of the pack. "Silicon Deepseek Reasoner" calls deepseek-ai/DeepSeek-R1 on SiliconFlow's API and splits the response into two strings: reasoning (the model's chain of thought) and answer (the actual reply).

It's the natural pairing to the pack's official-API reasoner node - same idea, but SiliconFlow's route, which matters if you're using a SiliconFlow account or free-trial quota instead of a platform.deepseek.com key. And it fixes one real gap in its DeepSeek-API cousin: this node does take a system_prompt.

How it works

Under the hood it's a plain HTTP POST to {base_url}/chat/completions with a bearer token, exactly like the Silicon chat node. The model is hardcoded to deepseek-ai/DeepSeek-R1. The reasoning is pulled from the response's reasoning_content field, and if the provider doesn't return one, you get the Chinese fallback "未提供推理过程" ("reasoning not provided") - worth knowing so that string doesn't spook you.

Like the other reasoner node, it keeps the conversation in memory on the node instance and replays it every run. Flip clear_history to true (or restart ComfyUI) to wipe it.

The inputs that matter

• prompt (multiline) - your question.
• system_prompt (default "You are a helpful assistant that can reason step by step") - required here, unlike the DeepSeek-API reasoner node. This is the one to set if you want the model to behave a certain way.
• clear_history (boolean, default false) - required. Toggle to true for a fresh conversation.
• temperature (0.7), max_tokens (default 512), top_p (0.7), top_k (50), frequency_penalty (0.5) - the SiliconFlow dial set. top_k is the one DeepSeek's own API lacks.

Outputs: reasoning and answer, both STRING.

Install and gotchas

Same pack, same routine: search "Comfyui-Deepseek" in ComfyUI Manager, or git clone https://github.com/yichengup/Comfyui-Deepseek into ComfyUI/custom_nodes, then restart. No model downloads - it's all API calls.

Key goes in custom_nodes/Comfyui-Deepseek/config.json under silicon_api_key (not api_key - wrong field means a Chinese error about the missing key). Get one from cloud.siliconflow.cn, edit, restart.

Real things to watch for:

• Low default max_tokens (512). R1 reasons before it answers, and 512 tokens goes fast. If answers come back cut off or the reasoning eats the whole budget, raise it toward 4096.
• Chinese error strings. 错误: or API请求错误 at the start of an output = a failed request returned as text (bad key, quota, network), not a shy model.
• History accumulates. In-memory, per node instance, and it keeps the assistant's answers - so long sessions get heavier on every call. clear_history is your reset button.
• "未提供推理过程" in the reasoning output just means the provider didn't return a trace this time. The answer output is still valid.

The verdict: of all five nodes in this pack, this one packs the most capability - two outputs, system prompt, history, and top-k tuning. It's the one to grab when you want R1 doing real work in a workflow, not just a one-off reply.