⚗️🔬| Load Qwen3-4B Model (safetensors / gguf)

Z-Image doesn't use a CLIP-L or T5-XXL text encoder like SDXL or Flux. It uses Qwen3-4B - a 4-billion-parameter bilingual LLM that's one of the two big reasons Z-Image understands natural-language and Chinese prompts as well as it does. This node is the pack's loader for that text encoder, and its selling point over the generic CLIPLoader is that it knows where to look for GGUF-quantized versions of Qwen3-4B too.

Load Qwen3-4B Model (safetensors / gguf) is one of the ⚗️🔬 experimental nodes, and it's effectively a convenience wrapper: you still need to download the actual encoder file and drop it in the right folder. What it saves you is the guesswork about which loader, format, and folder to use.

How it works

The node scans your text_encoders directory for Qwen3-4B checkpoints in .safetensors format, and - if you have city96's ComfyUI-GGUF custom nodes installed - it also finds .gguf files. You pick one from the checkpoint dropdown, and it loads the model as a CLIP using the lumina2 type that Z-Image expects. There's a file_filter toggle: off (default) shows only compatible Qwen3-4B files; on shows everything in the folder, which the tooltip admits includes files that may not be compatible.

The third input, run_on_gpu (default on), controls where the encoder runs. The tooltip wording is famously confusing here ("if enabled, executed on the CPU"), but the code is unambiguous: enabled = let ComfyUI place it on the default device (your GPU), disabled = force it to CPU. Flip it off when you're squeezing VRAM, and expect noticeably slower encoding - the CPU runs a 4B model about as well as you'd imagine.

Inputs & outputs

• checkpoint (COMBO) - pick your Qwen3-4B file.
• file_filter (BOOL) - show only compatible files vs everything.
• run_on_gpu (BOOL) - GPU (on) or CPU (off).
• Output: CLIP - wire into Style & Prompt Encoder or any text encoder.

Where the file goes: ComfyUI/models/text_encoders/. The README's recommended picks are qwen3-4b_int8_convrot_fp16emixed.safetensors (4.42 GB) alongside the ConvRot diffusion model, or Qwen3-4B-Q8_0.gguf (4.28 GB) if you're going GGUF - Q8 is basically fp16 at half the size.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/martin-rizzo/ComfyUI-ZImagePowerNodes

Restart ComfyUI. Two prerequisites: the Qwen3-4B file itself (that's the 4+ GB download - the node doesn't fetch it for you), and for GGUF specifically, city96's ComfyUI-GGUF must be installed or .gguf files won't even show up in the list.

Worth it?

If you already know your way around CLIPLoader with lumina2, this node is mostly a labeled shortcut. Where it genuinely earns its keep: the GGUF detection (saves you fighting ComfyUI's finicky handling of GGUF text encoders) and the one-click CPU fallback for low-VRAM rigs. It's experimental, so the usual caveat applies - the API could shift between versions. If your encoder won't load, the first things to check are the file being in text_encoders/ and, for GGUF, the ComfyUI-GGUF install.