LLM Siliconflow ♾️Mixlab

LLM Siliconflow is Mixlab's node for calling SiliconFlow's hosted API - a Chinese AI-infrastructure provider that, per the pack's own changelog, offers this specifically because it gives you a free LLM to use for prompt work without spinning up anything locally. Functionally it's a narrower cousin of the pack's ChatGPTOpenAI node: both are chat-completion clients you'd typically use to expand a rough idea into a detailed image prompt, translate, or summarize text inside your graph. The difference is ChatGPTOpenAI is a generic OpenAI-format client you can point anywhere, including a local server with no key at all, while this node is locked to SiliconFlow specifically and always needs an API key - there's no keyless local-endpoint escape hatch here.

The required fields are api_key (SiliconFlow issues these on siliconflow.cn; free-tier access is the whole point of this node existing), prompt (your actual message), and system_content, which defaults to a generic "answer as concisely as possible" instruction - replace it with something like "expand this into a detailed image-generation prompt" if that's what you're using the node for. model is a five-way dropdown of open models SiliconFlow hosts: Qwen/Qwen2.5-7B-Instruct (the default and a solid general pick), Qwen/Qwen2-7B-Instruct, THUDM/glm-4-9b-chat, 01-ai/Yi-1.5-9B-Chat-16K, and meta-llama/Meta-Llama-3.1-8B-Instruct - an optional custom_model_name field lets you target a different model SiliconFlow hosts that isn't in that shortlist. context_size (0–30, default 1) controls how many previous exchanges the node keeps in memory for multi-turn use; leave it low for one-shot prompt expansion. max_tokens runs from 512 up to 200,000 (default 512) - note the minimum is also 512, so you can't dial it lower even for a short reply. seed is there for reproducibility, standard for LLM sampling.

Outputs are text (the reply, what you'll wire onward most of the time), messages, and session_history - the latter two useful if you're chaining multi-turn conversations or want to inspect what actually got sent.

Install is the pack-wide standard:

cd ComfyUI/custom_nodes
git clone https://github.com/shadowcz007/comfyui-mixlab-nodes.git
cd comfyui-mixlab-nodes
install.bat

or via ComfyUI Manager (search "comfyui-mixlab-nodes"), or pip3 install -r requirements.txt in a venv. No local model download - this node is a thin API client, so the entire setup cost is getting an api_key from SiliconFlow's own dashboard.

The failure mode you'll hit most is auth: a missing or invalid api_key fails the call outright, and since this node has no local fallback the way ChatGPTOpenAI does, there's no way to keep working offline if that's wrong. If a request errors, check the key is current before assuming the node or model choice is at fault - free-tier keys and rate limits on hosted providers do change over time, so a key or model that worked last month isn't guaranteed to keep working without checking SiliconFlow's own dashboard. If you picked a model outside the five listed via custom_model_name and get a not-found error, double-check the exact model identifier against what SiliconFlow currently hosts - their catalog isn't fixed, and a typo or a retired model name will fail the same way a wrong key does. And if you're chaining multi-turn conversations and the model seems to "forget" earlier context, check context_size - it's the one field controlling how much history actually gets sent, and it caps out at 30.