license: gemma
base_model:
- Jiunsong/supergemma4-26b-abliterated-multimodal
language: - en
- ko
tags: - GGUF
- llama.cpp
- gemma4
- quantization
- turboquant
- tq3_4s
- multimodal
- abliterated
- conversational
pipeline_tag: image-text-to-text
Gemma4-26b-Super-Abliterated-TQ3_4S
Gemma4-26b-Super-Abliterated-TQ3_4S is a GGUF quantization of Jiunsong/supergemma4-26b-abliterated-multimodal using TQ3_4S, a 4.0 bpw Walsh-Hadamard-transform weight format with four per-8 scales.
Files
Gemma4-26b-Super-Abliterated-TQ3_4S.gguf(12.0 GiB)mmproj-f16.gguf(1.2 GiB)chat_template.jinja
Runtime Requirement
This model requires the public TurboQuant runtime fork:
It will not load correctly on stock llama.cpp or other runtimes that do not include TQ3_4S.
Text-Only Run
./build/bin/llama-server \
-m /path/to/Gemma4-26b-Super-Abliterated-TQ3_4S.gguf \
-ngl 99 -c 4096 -np 1 \
-ctk q8_0 -ctv q8_0 -fa on \
--cache-ram 0 --no-warmup --jinja \
--chat-template-file /path/to/chat_template.jinja
Vision / Image Input
./build/bin/llama-server \
-m /path/to/Gemma4-26b-Super-Abliterated-TQ3_4S.gguf \
-mm /path/to/mmproj-f16.gguf \
-ngl 99 -c 4096 -np 1 \
-ctk q8_0 -ctv q8_0 -fa on \
--cache-ram 0 --no-warmup --jinja \
--chat-template-file /path/to/chat_template.jinja \
--no-mmproj-offload
Performance (RTX 5060 Ti 16GB)
| Metric | Value |
|---|---|
| PP512 | 2154 tok/s |
| TG128 | 91.3 tok/s |
| Size | 12.01 GiB |
| BPW | 4.09 |
Quality
10/10 correct on standard QA benchmark (capital of France, 2+2, Python reverse string, gravity, WW2, primes, boiling point, Shakespeare, Jupiter, hello→Hola).
Base Model
Jiunsong/supergemma4-26b-abliterated-multimodal— abliterated multimodal variant of Google Gemma 4 26B-A4B- Original:
google/gemma-4-26B-A4B-it
Credits
- Jiunsong for the SuperGemma4 abliterated multimodal model
- Google DeepMind for Gemma 4
- huihui-ai for the abliteration technique
License
Same license terms as the base model apply (Gemma license).