SiliconFlow Service Connector 🦉| OwlV

SiliconFlow is a Chinese cloud that hosts a buffet of open models - DeepSeek, Qwen, GLM, Kimi - behind one OpenAI-compatible endpoint. This node is the menu. If you're in mainland China and don't want to juggle five provider accounts to translate your prompts, this is the connector you actually want: one key, many models, no proxy required.

Why you'd reach for it

The model dropdown is the whole pitch. deepseek-ai/DeepSeek-V3 gets you a frontier-class model without creating a separate DeepSeek account. Qwen/Qwen3-8B is a solid cheap workhorse for bulk translation. moonshotai/Kimi-K2-Instruct for something a bit different. And unlike the Gemini connector in this same pack, SiliconFlow works on a mainland network with no VPN - the README is explicit that domestic services are the intended path there.

How it works

The node is the pack's OpenAI-compatible connector pointed at https://api.siliconflow.cn/v1/chat/completions, with a provider-flavored payload - max_tokens 512, temperature 0.7, top_k 50, frequency_penalty 0.5. Standard Bearer auth with your key.

And yes, the same pattern as every connector here: running the node calls nothing. You get an llm handle out of it, and the real request fires when the LLM Translator node runs. If you run the connector alone and see no text, that's not a bug.

Inputs and output

• api_key - password-masked in the UI.
• model_select - six entries: deepseek-ai/DeepSeek-V3, THUDM/GLM-Z1-9B-0414, THUDM/GLM-4-32B-0414, Qwen/Qwen3-8B, moonshotai/Kimi-K2-Instruct, and Custom. Default is the GLM-4-32B snapshot.
• custom_model_id - the full model ID when you pick Custom.

Output is a single llm (type LLM) that plugs into the LLM Translator node.

Installing it

ComfyUI Manager: search ComfyUI-Translator, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/OwlvChirotha/ComfyUI-Translator
pip install -r ComfyUI-Translator/requirements.txt

Then restart. Dependencies are light - just requests (the other file in requirements.txt, googletrans, belongs to the pack's Basic Translator). No model downloads; everything is API calls.

Common gotchas

• The model list is pinned at author time, and SiliconFlow rotates its catalog. The default THUDM/GLM-4-32B-0414 is a dated snapshot. If a listed model 404s or has been deprecated, pick Custom and paste the current model ID from SiliconFlow's model page. Treat the dropdown as convenience, not gospel.
• 512-token output cap. Long text gets cut. Feed the translator reasonably sized chunks.
• Errors come back as text. A dead model or bad key surfaces as an error string in the LLM Translator's output rather than a crash - read the output.
• Your key is in the workflow JSON. The UI masks it; the saved file doesn't. The README tells you to clear keys before sharing workflows. Do that.

The pack itself is a small one-commit GPL-3.0 project with no real community footprint, but SiliconFlow is a real service people do use, and this node is a clean way to ride it inside ComfyUI. If you're on a China-friendly budget and want choice without account sprawl, start here.