license: gemma base_model: google/gemma-4-12B-it tags:
- abliterated
- uncensored
- gguf
- thai
- vision
- multimodal language:
- th
- en pipeline_tag: image-text-to-text
Gemma 4 12B Uncensored (Thai + English, Vision)
An abliterated (refusal-direction-removed) derivative of google/gemma-4-12B-it, with vision and Thai fluency.
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 degrades output quality in other languages. This model was abliterated using a Thai + English parallel prompt dataset, so it stays fluent in Thai while removing over-refusals. Vision is fully preserved — abliteration only touches the text side of the model.
Files
| File | Size | Use case |
|---|---|---|
gemma-4-12B-it-uncensored-beta1-Q4_K_M.gguf | ~6.9 GB | Main LLM, general use |
gemma-4-12B-it-uncensored-beta1-F16.gguf | ~22.2 GB | Main LLM, full precision |
gemma-4-12B-it-uncensored-beta1-mmproj-f16.gguf | ~167 MB | Vision projector (load with the LLM for image/audio input) |
Usage
Text (llama.cpp)
llama-cli -m gemma-4-12B-it-uncensored-beta1-Q4_K_M.gguf -ngl 99 --temp 0.6 --repeat-penalty 1.3 --repeat-last-n 512
Do not drop the temperature below ~0.6 (abliterated models loop at low temp).
Vision (llama.cpp)
Gemma 4 needs --jinja for its chat template, and a large batch size so the
image tokens fit in one non-causal pass:
llama-mtmd-cli -m gemma-4-12B-it-uncensored-beta1-Q4_K_M.gguf --mmproj gemma-4-12B-it-uncensored-beta1-mmproj-f16.gguf --image image.jpg -p "Describe this image" -ngl 99 --temp 0.6 --repeat-penalty 1.3 --jinja -b 4096 -ub 4096
The mmproj also supports audio input (experimental) via --audio.
LM Studio
Place the mmproj file in the same folder as the LLM; it is detected automatically and an image-attach button appears in chat.
How it was made
- Base model: google/gemma-4-12B-it (dense, multimodal)
- Method: abliteration (orthogonalizing the refusal direction out of the weights)
- Prompt dataset: Thai + English parallel prompts (harmful / harmless pairs). Using Thai prompts in the measurement step keeps the model's Thai output natural.
- Tool: jim-plus/llm-abliteration
License & responsible use
Use is subject to the Gemma Terms of Use. Removing refusals does not remove your responsibility — the prohibited-use policy still applies. Do not use this model for illegal or harmful activity.