Silicon Deepseek Chat

Same DeepSeek V3, different front door. "Silicon Deepseek Chat" talks to SiliconFlow's API (api.siliconflow.cn) instead of DeepSeek's own, which matters if you're outside DeepSeek's sign-up flow or just want SiliconFlow's free-trial quota. The pack author added these two Silicon nodes in February 2025 precisely so people could use DeepSeek models through SiliconFlow's account and billing instead of opening a separate platform.deepseek.com account.

The model is the same one the plain Deepseek Chat node uses (deepseek-ai/DeepSeek-V3), and the output is the same: one STRING with the answer. What changes is the plumbing and the tuning options.

How it works

Mechanically it's simpler under the hood: instead of the OpenAI SDK, the node makes a plain requests POST to {base_url}/chat/completions with a bearer-token header. No streaming, no fancy client - just a call and a parsed JSON reply.

The config is the interesting difference. Your key goes in the same config.json, but under silicon_api_key, not api_key - and if you fill the wrong field, the node returns a Chinese error telling you exactly that ("请在config.json中配置silicon_api_key"). Register at cloud.siliconflow.cn for a key.

The inputs that matter

prompt and system_prompt (required, same as the DeepSeek-flavored nodes). The optional set is where SiliconFlow flexes:

• temperature (default 0.7) - creativity, same as everywhere.
• max_tokens (default 512, cap 4096) - note the low default. SiliconFlow's free tier historically had tight per-request caps, so raise this if your answers get chopped.
• top_p (default 0.7) - sampling diversity.
• top_k (default 50) - this is the differentiator. SiliconFlow supports top_k (keep the K highest-probability tokens); DeepSeek's own API doesn't, which is why none of the non-Silicon nodes in this pack have it.
• frequency_penalty (default 0.5) - discourages repeated words.
• stop_sequence - end the reply the moment this string appears, handy for parseable output.

One output: STRING.

Install and gotchas

Standard pack install: search "Comfyui-Deepseek" in ComfyUI Manager, or git clone https://github.com/yichengup/Comfyui-Deepseek into ComfyUI/custom_nodes, then restart. Deps are openai and requests; nothing to download.

Edit custom_nodes/Comfyui-Deepseek/config.json:

{
  "api_key": "sk-deepseek-platform-key",
  "silicon_api_key": "sk-siliconflow-key"
}

Then restart ComfyUI - config is read when the node is created.

The trap to remember: this node's errors come back in Chinese. If the output starts with 错误: or API请求错误, that's the request failing (bad key, quota, network), returned as text. Don't assume the model got shy. And like the plain chat node, there's no conversation memory - every run is single-shot. If you want SiliconFlow-hosted R1 with history and a system prompt, the pack's Silicon Deepseek Reasoner is the other half of this pair.