gemma_3_12B_it_fp4_mixed.safetensors

Download gemma_3_12B_it_fp4_mixed.safetensors (9.4 GB)

Where to put it

Put the file in ComfyUI/models/text_encoders/ (in the portable version: ComfyUI_windows_portable/ComfyUI/models/text_encoders/). Older guides say models/clip: that folder still works. Then press R in ComfyUI, or restart it, so the file shows up in a text-encoder loader: Load CLIP (CLIPLoader) for one encoder, DualCLIPLoader when the model takes two (FLUX.1: CLIP-L + T5-XXL).

If a workflow says “Value not in list” for this node, the file name in the workflow is different from yours, or the file is in another folder. Click the node and pick the file from the list.

What it does

A text encoder turns your prompt into numbers the model understands. It runs once per prompt, then ComfyUI can move it out of VRAM to make room for the image model — so it costs load time and system RAM more than speed.

This is a smaller FP4 version. It loads faster and needs about half the RAM of the 16-bit file; the difference in the pictures is usually hard to see. It is the version this site lists by default.

FP4 needs an RTX 50 (Blackwell) card to run at full speed; on older cards take another version.

Models that use gemma_3_12B_it_fp4_mixed.safetensors

VRAM of the graphics card, for the model with the file this site picks for that card. Open a model to see every file and every GPU.

Other versions of this text encoder

Same encoder, different precision. Any of them works in the same loader node; pick one and select it in the node.

Not sure your card can run these models? Check your GPU. All shared files: ComfyUI model files.

Size and source read from Hugging Face (2026-10-01).