SLM: Model Selector

SLMModelSelector ("SLM: Model Selector") is the loader half of the pack's text-model side. You pick a model, a device and a dtype, and it hands the SLM generator a ready-to-use SLM_MODEL object plus the local path where the model lives. The generator doesn't care where the weights came from - it just consumes that object.

Right now there's exactly one choice in the model dropdown: Qwen3-0.6B, the 0.6-billion-parameter member of the Qwen3 family. That's a genuinely small model - the kind that runs comfortably on CPU-adjacent hardware and in a few gigabytes of VRAM. Calling the node a "selector" is a little generous when the menu has one item, but the pattern is the point: add entries to the pack's model table and this node picks them up.

How it works

The node does three things on execution:

1. Resolves the path. Models download into the pack's own Models/SLM/ folder (inside the custom node directory), named with the HF repo slug's / swapped for __ - so Qwen/Qwen3-0.6B becomes Models/SLM/Qwen__Qwen3-0.6B.
2. Downloads if missing. If the folder is empty and download_if_missing is on (it is by default), it pulls the repo from Hugging Face via snapshot_download. That first run can take a while depending on your connection.
3. Resolves device and dtype into the model dict. device auto-picks cuda when a GPU is available, else cpu. dtype auto becomes bfloat16 on a bf16-capable GPU, float16 otherwise, and float32 on CPU.

The inputs that matter

• model - the enum; currently just Qwen3-0.6B.
• device - auto (default), cuda, or cpu.
• dtype - auto (default), bfloat16, float16, or float32.
• download_if_missing - on by default. Turn it off and the node errors instead of downloading when weights are absent.

Two outputs: slm_model (the SLM_MODEL object) and model_path (the local folder as a STRING, handy for a display node or if you want to know where the gigabytes went).

Install

Part of the comfyui_pilcothink_VisionSLM pack:

cd ComfyUI/custom_nodes
git clone https://github.com/gpdev-Pilcothink/comfyui_pilcothink_VisionSLM

Restart ComfyUI, or install via ComfyUI Manager (search comfyui_pilcothink_VisionSLM). The pack needs transformers ≥ 4.56.2 for Qwen3 support, plus torch and accelerate from its requirements.

Gotchas

The big one is where the weights land: inside the custom node folder, not ComfyUI/models/. Uninstalling or updating the pack can take your downloaded models with it. If you want to move them, you can - the folder layout is self-contained, so a symlink from Models/SLM to a big disk works.

Two more from the README. If you pick cpu as the device, the author says it only works with float16; other dtypes error out, so don't fight it, just set it. And the README's "downloads into Models/SLM_Vision/" is stale - the code actually uses Models/SLM for this node and Models/VisionSLM for the vision one. Trust the code. As with the rest of this pack, it's new with no community track record, so grab it from the official repo only.