小珠光 Qwen Model Loader

This node loads a local LLM into your graph so other Xiaozhuguang nodes can use it - specifically the MiniMax H3 prompt formatter, which needs a BSAI_QWEN_MODEL on its input. XiaozhuguangQwenModelLoader (小珠光 Qwen Model Loader) is the pack's own way to produce that model handle: pick a GGUF file from ComfyUI/models/LLM/, and out comes a qwen_model you can wire into the H3 node. The name is a little misleading - it's a loader for a family of Qwen/Gemma GGUF models, not a Qwen-specific inference node.

Here's the honest framing before you get excited: this node does not run Qwen-VL inference. That's a different node in the same pack (XiaozhuguangQwenVLInstruct) with its own transformer-based loader and its own model folder. This one is the GGUF path - the quantized-weights format that lets a chat LLM share a consumer GPU with your diffusion model, which is exactly the VRAM-budgeting story the KB's GGUF panel tells. The node scans ComfyUI/models/LLM/, lists the main .gguf/.safetensors/.bin files as model_file, and lists files with "mmproj" in the name (the multimodal projector) separately as mmproj.

Inputs worth caring about:

• model_family - Qwen3-VL, Qwen3.5-VL, Qwen3.6-VL, or Gemma4. This is metadata that steers prompt formatting downstream - pick the family that matches the GGUF you actually loaded.
• model_file - the main model. Drop a GGUF into ComfyUI/models/LLM/ and restart, and it appears here. The placeholder "(将模型放入 models/LLM)" means the folder's empty.
• mmproj - set this when your model is a vision model and you want images handled (the H3 node's image_1…9 inputs only mean anything if a projector is loaded). Pure text? Leave it None.
• context_length - the KV-cache context window, default 32768. This is the VRAM knob: context length and cache memory scale together, so if you're OOMing, drop this before you touch anything else.
• gpu_layers - how many layers go on the GPU; -1 (default) puts everything on GPU and auto-degrades on low VRAM. If you're sharing the card with a video model, this is where you compromise.

One output: qwen_model - the loaded handle, for the H3 node (or anything else that speaks BSAI_QWEN_MODEL).

The heavy dependencies here are real: this node needs llama-cpp-python for GGUF loading, which is in the pack's declared dependency list (registry install handles it; a git install means running the pack's requirements). And it needs the model files themselves, which you supply - the pack won't download them. Given the KB's GGUF guidance: Q8 is basically fp16 at half the size, Q5 is the last stop before visible loss, and for a prompt-rewriting LLM you can absolutely get away with a Q4 - the model's job here is short structured rewriting, not genius.

Install the pack via ComfyUI Manager (ComfyUI-xiaozhuguang) or git clone, restart, and put your GGUF in ComfyUI/models/LLM/. Pair this with the H3 node, load a quant, and you've got a fully local prompt-formatting pipeline - no API key, no cloud, and nothing phoning home.