In ComfyUI v0.15.0 they added a new native "TextGenerate" node: Comfy-Org/ComfyUI#12392
It works, but extremely slow on my 12gb card due to offloading. Gguf would be very useful. Currently gguf doesn't work - I load gemma 3 12b qat Q4_0 model in "CLIPLoader (GGUF)" node, type ltxv, and get error RuntimeError: mat1 and mat2 shapes cannot be multiplied (4096x1152 and 4304x1152)