🌐 Remote Text Model Config (Ollama/Nexa/LM Studio)

Not everyone wants to fight llama-cpp-python for local inference. Some days you want your LLM running in a tool that already handles model downloads, GPU config, and a server - and ComfyUI just talks to it over HTTP. This node is the remote half of the ComfyUI-GGUF-VLM pack's text story: it turns any Ollama, LM Studio, or Nexa SDK instance into a model source your graph can use, all through OpenAI-compatible endpoints.

How it works

The mechanism is blessedly simple. Pick an api_type, point it at a base_url, and the node hits {base_url}/v1/models to populate the model dropdown. Since Ollama, LM Studio, and Nexa all speak OpenAI-compatible formats on their local ports, one config node covers all three:

• Ollama - default http://127.0.0.1:11434
• LM Studio - http://127.0.0.1:1234 (the README's recommended path for Windows, since it bundles its own Python/CUDA and downloads GGUF models through a GUI)
• Nexa SDK - http://127.0.0.1:8080

model is the dropdown, populated by clicking the 🔄 Refresh Models button once your server is running and a model is loaded. There's an optional system_prompt field for a default persona. Outputs are model_config (TEXT_MODEL), which feeds the pack's remote text generation node, plus status_info (STRING) showing what the connection reported.

The inputs a beginner actually touches

Three things, in this order: start your service (ollama serve, or launch the LM Studio local server), set api_type to match, and make sure base_url has the right port. That's it. The default ports in the tooltips are the ground truth - LM Studio is 1234, Ollama 11434, Nexa 8080 - so the most common failure is just having the wrong port or the service not running.

Install and gotchas

Installation is the standard pack dance:

cd ComfyUI/custom_nodes
git clone https://github.com/walke2019/ComfyUI-GGUF-VLM.git
cd ComfyUI-GGUF-VLM
pip install -r requirements.txt

Note you do not need the CUDA llama-cpp-python build for the remote path - the heavy inference happens in the external server, not ComfyUI. That's the whole appeal: on a machine where you can't or won't compile llama-cpp with CUDA, remote mode gets you local LLMs anyway.

The traps are all about the server side. If the model dropdown stays empty, the service isn't reachable or isn't serving a model yet. If your generation is slow or times out, remember the vision sibling of this node defaults to a 300-second timeout for a reason - local VLMs are not fast. And the "OpenAI Compatible" option that shows up on the vision config isn't here: remote text is exactly Ollama, Nexa, or LM Studio, so don't hunt for a generic endpoint that this node doesn't expose.