Local AI Model

Local AI Model is the on-ramp to the whole Local AI half of the WepeNerd pack. You drop it on the canvas, pick a GGUF model from the dropdown, and wire its single model output into Prompt Enhancer, Image Captioner, or Video Captioner. That's the whole normal workflow - model in, text out, no API keys, no cloud, no per-call cost. If you've ever wanted an uncensored, offline prompt rewriter or captioner sitting inside your graph, this is the node that makes it a four-node pipeline instead of a project.

Here's the thing that trips people up: this node does not load the model into ComfyUI's memory. It builds a config object (GGUF_LLM_CONFIG) that tells the pack to spawn a real llama-server process - llama.cpp's HTTP server - on localhost the moment a downstream generation node runs. ComfyUI never sees the weights as a Comfy model, which is both the clever part and the source of the VRAM confusion covered later.

The two inputs that matter:

• model - a dropdown populated from .gguf files in ComfyUI/models/LLM (subfolders are fine). The pack registers the folder itself, so drop files there and refresh. The README's reference setup is an abliterated Qwen3 GGUF like Huihui-Qwen3.8-27B-abliterated-Q4_K.gguf. If the dropdown only shows <put .gguf models in ComfyUI/models/LLM>, that's the pack telling you it found nothing - and running will error, not silently use a random file.
• projector - Auto / None by default. Pick a vision projector (something with mmproj in the name, like mmproj-model-bf16.gguf) only when you're captioning images or video. The pack tries to separate models from projectors by filename token, so keep mmproj/vision in the filename and it'll sort itself.

What you get is the pack's safe default set baked in: 8192-token context, all layers offloaded (gpu_layers = -1), it asks ComfyUI to free about 24 GB of VRAM before starting the server, and it releases the server after every generation so VRAM comes back. Local AI Model (Advanced) is the same node with every knob exposed; use this one until you know exactly why you need the advanced version.

Installing. Via ComfyUI Manager, search for WepeNerd and install. Or:

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

Then restart. One honest gotcha: the pack's requirements.txt pulls in the 3D render stack too (trimesh, pyrender, PyOpenGL, pyglet), so the install is heavier than the Local AI nodes strictly need - it's a shared install for the whole pack.

The one thing the pack will not download for you is llama.cpp itself. These nodes talk to a real llama-server executable, and it has to exist before anything runs. Put it on PATH, or set LLAMA_SERVER_PATH, or extract a Windows build to C:\llamacpp\, or paste its full path into the advanced config. A .gguf file is not an executable - that's the #1 "why does nothing happen" failure. If startup fails, the ComfyUI console shows the tail of llama-server's own output; read that before assuming the node is broken.