Lockout

/

qwen3-4b-heretic-zimage

like 11

I also also looking forwards to the safetensors as gguf isn't supported in many places. Also, full precision would be great. It doesn't have many parameters and the full model can fit into VRAM without a problem.

Ok, I can give you what I got (BF16) but my upload is shit. I also wanted to experiment with larger dataset and the grimjim ablation. Same with picking the next one down of 3 refusals but much lower KLD. Full model is 8gb, on my internet that's I think a 2hr upload or more.

RESOLVED BY UPDATING THE GGUF NODES.

Gives an error in ComfyUI. Is this supposed to be a Clip Text Encoder?

Works with the GGUF clip text encoder node but it's an LLM which passes embeddings to DIT.