LLMs Loader

The config node you only need when you want the explicit path

LLMs Loader (class LLM_Loader) is the odd one out in ComfyUI-LLMs-Toolkit: it doesn't call any API, it doesn't generate anything, and it needs no key of its own at all. What it does is package a provider, model, and API key into a single LLM_CONFIG value that you wire into other nodes in the pack - most usefully the OpenAI Compatible Adapter and LLM Translator. Think of it as the manual gearbox for a pack whose default is automatic: if you want your model configuration visible on the canvas instead of tucked away in the LLMs_Manager settings panel, this is how you get it.

The catch is that most people don't need it. The pack's normal flow is: set up providers once in the LLMs_Manager UI, then let the adapter's dropdowns do the work. This node exists for the cases where that flow is the wrong tool - a workflow you share with someone whose provider setup differs from yours, a quick test against a throwaway key, or a custom endpoint that isn't in the preconfigured list.

How it works

You pick a provider from the dropdown - 12 preconfigured Chinese API providers (Qwen/通义千问, DeepSeek/深度求索, GLM/智谱清言, Moonshot/月之暗面, DouBao/豆包, Spark/星火, Baichuan/百川, MiniMax, StepFun, SenseChat, iFlow, ModelScope) plus Custom/自定义 - type a model name and an api_key, and it returns an llm_config bundle. Behind the scenes each built-in provider has a hardcoded base URL (e.g., DeepSeek → https://api.deepseek.com/v1, Qwen → https://dashscope.aliyuncs.com/compatible-mode/v1), so you only ever supply the model string and key.

If you pick Custom/自定义, the required custom_base_url input appears and becomes the endpoint - which is how you point this at anything OpenAI-compatible that isn't in the list. A local Ollama server at http://localhost:11434/v1 works exactly the same way. The node even validates before executing: custom provider without a base URL, or an empty model or key, and it refuses with a message instead of producing a half-broken config.

Inputs and output

• provider - the 13-way enum described above; the only field a beginner really sets.
• model - a plain string, e.g. deepseek-chat or qwen-plus. Case matters; it's passed straight to the API.
• api_key - paste your key. It gets baked into the llm_config bundle, so this is the one place a key can end up embedded in a saved workflow - be careful about sharing those files.
• custom_base_url (optional) - required in effect when provider is Custom.
• Output: llm_config (type LLM_CONFIG), a custom pack-wide type. Wire it into the adapter's llm_config input, and set the adapter's provider dropdown to "LLM_CONFIG (from input)".

Installing

It's part of the HuangYuChuh/ComfyUI-LLMs-Toolkit pack, so install once, get all twelve nodes. ComfyUI Manager → search ComfyUI-LLMs-Toolkit → Install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/HuangYuChuh/ComfyUI-LLMs-Toolkit
cd ComfyUI-LLMs-Toolkit
pip install -r requirements.txt

Restart afterwards. Dependencies are featherweight - Pillow and aiohttp - and there are no model downloads.

Gotchas

The honest take: unless you're sharing workflows, testing a second provider, or hitting a local model, you'll probably forget this node exists - and that's fine. When you do use it, the two traps are (1) leaving the provider dropdown as an API provider while also typing a custom base URL, and (2) assuming the LLM_CONFIG type is interchangeable with the adapter's provider dropdown. It isn't: the adapter's provider field must explicitly say "LLM_CONFIG (from input)" or the bundle you built gets ignored. And since the key rides inside the config, a workflow saved with this node contains your API key in the JSON - scrub it before you share.