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:

https://github.com/turbo-tan/llama.cpp-tq3

It will not load correctly on stock llama.cpp or other runtimes that do not include TQ3_4S.

Text-Only Run

bash

./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

bash

./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).

YTan2000/Gemma4-26b-Super-Abliterated-TQ3_4S

Gemma4-26b-Super-Abliterated-TQ3_4S