Siliconflow paid chat

Some prompts deserve a 72B model, and no amount of clever quantization is putting a 72B on your 8GB card. PaidChat-2lab doesn't try. It sends your text to SiliconFlow's paid LLM API and streams back an answer - Qwen2-72B, DeepSeek-V2, even Qwen1.5-110B, all running on their servers while your GPU stays free for diffusion. The trade is money instead of VRAM, and for a one-off prompt-expansion or a story-bible generation, that's often the right trade.

PaidChat (2lab) is the bigger sibling in AI2lab's comfyUI-siliconflow-api-2lab pack, the same tiny MIT wrapper that ships FreeChat. Same mechanism, different menu: where FreeChat lists the platform's free-tier models, this node's dropdown is the paid lineup - Qwen2-72B-Instruct (the default), the Qwen2-57B-A14B mixture-of-experts model, Qwen1.5-110B/32B/14B, DeepSeek-Coder-V2-Instruct, DeepSeek-V2-Chat, DeepSeek-LLM-67B, and Yi-1.5-34B-Chat-16K. These are the models you genuinely cannot run at home. You're renting them per token.

How it works. Identical plumbing to FreeChat: an OpenAI-style chat completion against https://api.siliconflow.cn/v1, your prompt as the user turn, a fixed "You are a helpful assistant" system message, stream=False, reply returned as a text STRING. Your API key is read from config.json in the pack folder when ComfyUI starts, so that file is your entire setup. The one real difference in daily use is latency - a 110B model under load can keep your queue blocked for tens of seconds. That's the nature of the sync call, not a bug.

Inputs that matter:

• prompt - multiline. This node earns its keep on long, structured asks: "Write a 200-word setting description for a cyberpunk detective story, in the style of a T5-friendly caption." Small talk is a waste of paid tokens.
• model - the paid dropdown above. The MoE Qwen2-57B-A14B is a nice middle ground (57B total, only ~14B active per token, so it's fast and smart). Default to Qwen2-72B unless you have a reason.

No temperature, max tokens, or seed are exposed. The other real limitation: the model list is hardcoded as an enum. It's frozen at 2024-era DeepSeek and Qwen - no DeepSeek-V3/R1, no newer Qwen, and you can't type a model that isn't listed. If the model you want isn't in the dropdown, this node is a dead end.

What the text output wires into. It's a plain STRING, so the standard moves apply: into a CLIPTextEncode text input to use the model's answer as your actual generation prompt, or into a text-display node to review before generating. Use it as the smart layer that turns your vague idea into the elaborate prompt the diffusion model then renders.

Install. ComfyUI Manager: search "comfyUI-siliconflow-api-2lab". Or:

cd ComfyUI/custom_nodes
git clone https://github.com/AI2lab/comfyUI-siliconflow-api-2lab

Only dependency is openai; there are no model files to download. After the first launch the pack creates config.json from its template - put a real API key from https://cloud.siliconflow.cn/ in it and restart ComfyUI. The key is read at import time, so restart is mandatory after editing.

Troubleshooting worth knowing:

• 401s mean the placeholder "your key" is still in config.json, or the key is bad. Replace, save, restart.
• Node missing after install = config.json is broken JSON or lacks API_KEY; the import dies and takes the node with it.
• Signup friction: the .cn console can demand a Chinese phone number for verification, which has tripped up non-Chinese users. The international console at siliconflow.com is the escape hatch.
• The bill. These are paid models, billed per token. A long story-bible generation on Qwen-110B is still cheap in absolute terms, but it's not zero - and free-tier quota does not apply to anything in this dropdown. That's literally why the pack has two chat nodes.

Since a node like this is a remote-API call by design, the security question gets asked: after the ComfyUI_LLMVISION malware incident, that skepticism is earned. This pack is ~90 lines of readable MIT code whose only outbound call is SiliconFlow's API with your key. Audit it in five minutes if you want the peace of mind - you won't find anything else.