inflatebot/LTX23-gemma-3-12b-it-orthogonal-reflection-bounded-ablation-v4-fp4_mixed

An fp4_mixed quantization of grimjim/gemma-3-12b-it-orthogonal-reflection-bounded-ablation-v4-12B, intended to generate LTX2(.3) embeddings and captions.

ORBA is very gentle, enough that even at fp4_mixed, it retains enough brainpower to generate captions.

But the GGUF CLIP loaders I was using weren't compatible with the TextGenerateLTX2Prompt node, so it kept spitting out tensor size mismatch errors.

So I contracted Opus 4.6 to analyze the structure to verify that it was a standard template (because I honest to god didn't know), and then it wrote a script to quantize and graft the ablated tensors, and after a hiccup with the quantization method (see below), ended up with something uncensored and capable that fits within 10GB.

The scripts Opus prepared are here too, but don't ask me how to use them, because for half of the session it may as well have been speaking in tongues. 😅

Vibe-Quanting. Not even once.