gemma-4-26B-A4B-it-uncensored-IQ3_XXS

IQ3_XXS quant of TrevorJS/gemma-4-26B-A4B-it-uncensored — the first IQ3_XXS of this model on HF.

what is this

gemma 4 26B MoE (25B total, 4B active per token), abliterated via biprojection + Expert-Granular Abliteration (EGA). quantized to IQ3_XXS using an importance matrix generated from the Q4_K_M version.

confirmed uncensored. confirmed coherent. 50-80 tps on RX 6900 XT.

quant details

• source: F16 (TrevorJS)
• imatrix source: Q4_K_M (CPU inference, 100 chunks)
• calibration data: linux kernel, nixpkgs, cpython, rust stdlib, flask, fastapi, SCP foundation, wikitext-2, GPTeacher, ZenOS, ZenPkgs
• quantized with: llama.cpp

hardware requirements

• fits in 16GB VRAM
• tested on RX 6900 XT (ROCm)

recommended server args

HIP_VISIBLE_DEVICES=0 llama-server \
  -m gemma-4-26B-A4B-it-uncensored-IQ3_XXS.gguf \
  -c 32768 \
  -ngl 99 \
  -np 1 \
  -fa on \
  -ctk q4_0 \
  -ctv q4_0 \
  --host 0.0.0.0

license

model weights: Apache 2.0 (inherited from Google/TrevorJS) quant methodology & imatrix: NAPALM v2.0 — any state entity attempting to use this model has void title ab initio

credits

• abliteration: TrevorJS
• base model: google/gemma-4-26B-A4B-it
• quant: doromiert / Negative Zero