Gemma 4 26B A4B Uncensored TH (GGUF)
An abliterated (refusal-direction-removed) derivative of google/gemma-4-26b-a4b-it.
This is a modified model derivative created under the Gemma Terms of Use.
What makes this different
Most abliterated models are measured with English-only refusal datasets, which often degrades output quality in other languages. This model was abliterated using a Thai prompt dataset (with English refusal markers kept as a fallback), so it stays fluent in Thai while reducing over-refusals.
This upload is text GGUF only. Vision/mmproj is not included here.
Files
| File | Quant | Size (approx.) | Use case |
|---|---|---|---|
gemma-4-26B-A4B-it-heretic-th-Q5_K_M.gguf | Q5_K_M | 16–18 GB | recommended |
gemma-4-26B-A4B-it-heretic-th-Q4_K_M.gguf | Q4_K_M | 13–15 GB | 32 GB VRAM |
gemma-4-26B-A4B-it-heretic-th-Q3_K_M.gguf | Q3_K_M | 11–13 GB | smaller footprint |
Usage
llama.cpp
./llama-cli \
-m gemma-4-26B-A4B-it-heretic-th-Q5_K_M.gguf \
-ngl 99 \
--reasoning off --reasoning-budget 0 \
-c 8192 \
--temp 0.7 \
--min-p 0.05 \
--top-p 0.9 \
--repeat-penalty 1.12 \
--repeat-last-n 256
Do not drop the temperature below ~0.6. Abliterated models tend to loop at very low temperature. For longer writing, --temp 0.78 and -c 16384 work better than raising --repeat-penalty.
LM Studio
Load the GGUF directly. Turn thinking/reasoning off if the UI exposes that toggle.
How it was made
Base model: google/gemma-4-26b-a4b-it (MoE, 26B total / 4B active) Method: abliteration via Heretic Prompt dataset: Thai harmful / harmless pairs used for measurement and evaluation, so Thai output stays natural after the edit Selected trial: 199 (0/16 refusals, KL 0.0074)