Jiunsong/supergemma4-26b-abliterated-multimodal-gguf-4bit

Support ongoing open-source work: ko-fi.com/jiunsong

SuperGemma4-26B-Abliterated-Multimodal GGUF 4bit

This is the compact llama.cpp-ready GGUF 4bit distribution of Jiunsong/supergemma4-26b-abliterated-multimodal.

It keeps the matching multimodal projector and was validated with both text and image prompts after quantization.

Included files

• supergemma4-26b-abliterated-multimodal-Q4_K_M.gguf
• mmproj-supergemma4-26b-abliterated-multimodal-f16.gguf

Validation

• Text check: returned READY
• Image check: returned Red for a solid red test image
• Text throughput in llama.cpp: prompt 230.6 tok/s, generation 137.1 tok/s
• Image throughput in llama.cpp: prompt 138.1 tok/s, generation 50.3 tok/s
• Disk footprint: about 17 GB

Quantization note

This build was generated as Q4_K_M. A small number of tensors were automatically kept at higher precision by llama.cpp where needed for compatibility and stability.

Recommended use

Use this build when you want the smallest practical GGUF package here while keeping text + vision capability.

Quick start

llama-cli \
  -m /absolute/path/to/supergemma4-26b-abliterated-multimodal-Q4_K_M.gguf \
  -mm /absolute/path/to/mmproj-supergemma4-26b-abliterated-multimodal-f16.gguf \
  -cnv -st \
  --image /absolute/path/to/image.png \
  -p "Describe the image briefly."