SuperGemma abliterated multimodal finetune converted to GGUF and quantized with an unofficial reapplication of the PRISM-PRO-DQ per-tensor type map. The multimodal projector is not bundled in this repository.

Important provenance notice

This is not an official PRISM-DQ or PRISM-PRO-DQ model. This repository is not affiliated with Ex0bit, PRISM-DQ, or the authors of the source finetune.

The underlying trained weights come from Jiunsong/supergemma4-26b-abliterated-multimodal. No trained tensor values from Ex0bit's PRISM model were copied, merged, or transplanted into this model.

The quantized file was produced by reconstructing the per-tensor GGML quantization-type assignment visible in the public Ex0bit/Gemma4-26B-A4B-PRISM-PRO-DQ-GGUF release and then applying that assignment to this separate finetune. This is therefore a map reapplication, not a reproduction of the original PRISM-DQ calibration, optimization procedure, dataset, or proprietary pipeline.

The Unsloth importance matrix imatrix_unsloth.gguf_file was used during quantization. Conversion and quantization were performed with llama.cpp.

Quantization layout

For the 658 tensors shared with the reference Gemma 4 PRISM-PRO-DQ GGUF, the final distribution is:

GGML type	Tensor count
F32	392
Q5_1	30
Q5_K	91
Q6_K	115
Q8_0	30

The language-model GGUF contains 658 tensors and matches the reconstructed reference type map without extra language-model tensors.

Files and integrity

File	Size	SHA-256	Purpose
`Gemma-4-26B-A4B-SuperGemma-Abliterated-Multimodal-PRISM-DQ-I.gguf`	16.85 GiB	`bff06c05e613b471da0dda3b801ad090c174d071c7299d707a98802b73dc17c3`	Unofficial PRISM-PRO-DQ tensor-map reapplication with Unsloth I-matrix

Template and tokenizer

The tokenizer vocabulary and special-token IDs were checked against the source Hugging Face tokenizer.
The embedded chat template was updated to the current official template from google/gemma-4-26B-A4B-it at conversion time.

Text-only usage with llama.cpp

bash

llama-cli \
  -m Gemma-4-26B-A4B-SuperGemma-Abliterated-Multimodal-PRISM-DQ-I.gguf \
  --jinja \
  -p "Hello"

For server use:

bash

llama-server \
  -m Gemma-4-26B-A4B-SuperGemma-Abliterated-Multimodal-PRISM-DQ-I.gguf \
  --jinja \
  --host 127.0.0.1 \
  --port 8080

Multimodal usage

This repository does not include an mmproj file. To enable image input, download a compatible Gemma 4 26B A4B multimodal projector from the Unsloth GGUF repository:

unsloth/gemma-4-26B-A4B-it-GGUF

Download one of the available mmproj files from that repository and pass its local path to llama-mtmd-cli:

llama-mtmd-cli   -m Gemma-4-26B-A4B-SuperGemma-Abliterated-Multimodal-PRISM-DQ-I.gguf   --mmproj path/to/mmproj.gguf   --image path/to/image.jpg   -p "Describe this image."

References

Source finetune: Jiunsong/supergemma4-26b-abliterated-multimodal
Official Gemma 4 base: google/gemma-4-26B-A4B-it
Unsloth GGUF and projector repository: unsloth/gemma-4-26B-A4B-it-GGUF
PRISM-PRO-DQ public GGUF used only as the tensor-map reference: Ex0bit/Gemma4-26B-A4B-PRISM-PRO-DQ-GGUF
Unsloth importance matrix: imatrix_unsloth.gguf_file
GGUF conversion and quantization tooling: ggml-org/llama.cpp
Gemma Terms of Use: https://ai.google.dev/gemma/terms

License and responsibility

Use of this repository remains subject to the original source model's license and the Gemma Terms of Use. The behavior and safety characteristics of the source finetune are inherited; quantization does not make the model safer or change its intended-use restrictions. Users are responsible for evaluating outputs and complying with applicable laws and licenses.

KaviarLabs/Gemma-4-26B-A4B-SuperGemma-Abliterated-Multimodal-PRISM-DQ-I

Gemma 4 26B A4B SuperGemma Abliterated Multimodal — PRISM-PRO-DQ Map Reapplication