CXH_MinCP3_4B_Load

CXH_MinCP3_4B_Load loads MiniCPM3-4B, the text-only language model in this pack. Note the distinction from the MiniCPM-V nodes: MiniCPM-V 2.6 sees images, MiniCPM3-4B is pure text. The README's changelog calls it out for "chat, translation, and rewriting" - a general-purpose local LLM you can drop in a ComfyUI graph, not a captioner.

It's the odd one out in a pack otherwise dedicated to vision, and honestly that's a good thing to know before you install: unless you have a specific text task in mind (translate a caption, rewrite a prompt, polish a dataset's text side), you may never touch this pair. But when you want a local, no-API-key chat model inside ComfyUI, it's convenient - one loader, one chat node, done.

What it loads

One input, one output:

• Input: model - two choices: openbmb/MiniCPM3-4B (full) or openbmb/MiniCPM3-4B-GPTQ-Int4 (4-bit quantized, for tighter VRAM). If your card has the headroom, the full model is the safer bet for quality; the Int4 exists precisely for 8 GB-ish setups.
• Output: pipe (CXH_MinCP3_4B_Pipe) - the model + tokenizer, loaded in bfloat16 with device_map="auto".

Downloads go to models/LLM/<model name> via the pack's standard helper on first run.

Installing

Same pack install - ComfyUI Manager (search "Comfyui_CXH_joy_caption") or:

cd ComfyUI/custom_nodes
git clone https://github.com/StartHua/Comfyui_CXH_joy_caption
cd Comfyui_CXH_joy_caption
python -m pip install -r requirements.txt

Gotchas

• 4B but not tiny: bfloat16 4B is still a few GB in VRAM, and it's loaded with device_map="auto" which will happily spread onto CPU if the GPU fills up - expect slow generation in that case.
• The GPTQ-Int4 option needs auto-gptq/optimum-style loading support in the transformers version you have. If that choice errors at load, it's an environment issue, not a model problem - the full model usually loads cleanly.
• Text-only. Don't go looking for an image input on the chat node; there isn't one. For vision+prompt work you want the MiniCPM-V 2.6 nodes (CXH_HG_Model_Load + CXH_Min2_6_prompt_Run).
• It's a general LLM, so the pack's own captioning nodes remain the right tool for datasets - this one is for when you need to do something with text.