Set SiliconFlow LLM Service Connector 🐑

SiliconFlow (硅基流动) isn't a lab - it's an inference host that runs other people's open-weight models cheaply, and that's exactly why it's one of two providers the README singles out as offering free-tier access (the other is ZhiPu). The catch, which the README is upfront about, is that the free lineup rotates: whatever's free today might not be next month, so you check SiliconFlow's own site for the current list and drop the model ID into custom_model rather than trusting the dropdown to stay current. This connector doesn't generate anything itself - same as every Set*LLMServiceConnector node in this pack, it packages your key and model choice into an LLMServiceConnector object for a downstream node (Translator, a prompt generator) to actually use.

Reading the model list. The default is deepseek-ai/DeepSeek-V4-Flash, and the other 15 choices are worth a second look because the naming isn't random - it's SiliconFlow's own catalog convention showing through. You'll see plain namespace/Model entries (deepseek-ai/DeepSeek-V3, Qwen/Qwen3.6-35B-A3B) alongside Pro/namespace/Model variants of the same models (Pro/deepseek-ai/DeepSeek-V3.2, Pro/zai-org/GLM-5.1, Pro/moonshotai/Kimi-K2.6) - the Pro/ prefix marks SiliconFlow's dedicated, faster-hosted tier of an otherwise-open model, distinct from the bare namespace which is typically the free or lower-priority-hosted route. And if you spot both zai-org/GLM-4.5V and THUDM/GLM-4-32B-0414 in the same list and wonder why one company shows up under two names: that's genuinely one company, not two. Zhipu AI rebranded to Z.ai in 2025, and its older model repositories still sit under the pre-rebrand THUDM GitHub org while newer ones publish under zai-org - SiliconFlow's catalog just reflects both eras verbatim.

Inputs that matter. api_token is your SiliconFlow key (sk-..., from cloud.siliconflow.cn). model_select covers the 16 choices above; custom_model overrides it with any model ID SiliconFlow currently hosts, which given how often their catalog and pricing shift is genuinely the field you'll reach for most.

The config-file gotcha. prefer_local_config defaults to true: the node reads mie_llm_keys.json (copied from mie_llm_keys.json.example, kept beside the README, deliberately not committed to git) and uses the siliconflow entry there ahead of whatever's typed into the node's own api_token. This is the single most common source of confusion across every connector in this pack - a key typed into the node that gets silently ignored because a stale one sits in the JSON file. Fix the file, or flip prefer_local_config to false to make the node's field win. config_key lets you read a different entry if you keep multiple SiliconFlow accounts.

Install. Search "ComfyUI_MieNodes" in ComfyUI Manager, or cd ComfyUI/custom_nodes && git clone https://github.com/MieMieeeee/ComfyUI-MieNodes and restart. No downloads - this is credential plumbing, not a hosted model.

Troubleshooting. If a model in the dropdown 404s or 400s, it's very likely been renamed, retired, or moved between the free and Pro/ tiers since this node's list was last updated - SiliconFlow's catalog changes often enough that this is closer to expected than exceptional. Check the current model list on cloud.siliconflow.cn and use custom_model with whatever ID is live now. If you were counting on a free-tier model specifically, verify it's still marked free before you burn a debugging session assuming the connector is broken - the pricing tier under a model name is exactly the kind of thing that moves without warning.