fredrezones55/ Gemma-4-Uncensored-HauhauCS-Aggressive:e4b-SCN

2,856 2 days ago

Bringing the monolithic GGUF to HauhauCS's work of aggressively uncensoring Gemma 4. For edge models having full audio and vision support; there are multi-modal specific focused models available [Vision OR Sound].

vision tools thinking audio e2b e4b 
ollama run fredrezones55/Gemma-4-Uncensored-HauhauCS-Aggressive:e4b-SCN

Details

2 days ago

4627d8e69e96 · 5.7GB ·

model
gemma4
·
7.69B
·
Q4_K_M
params
{ "temperature": 1, "top_k": 64, "top_p": 0.95 }
template
{{ .Prompt }}

Readme

models:

All E2B and E4B models have full image and audio support; the uncensoring does not seem to cause faults in the audio processing. This is not a STT model, but is to respond to focused tasks about the audio or image it sees. (Personally I prefer the E4B model for some audio tasks) [will come back and resolve issues with the E2B model later…]

As an experiment: I created partially gutted models for specific multimodal tasks

Multimodal focused models [🔊, 📷, 📝]

SND SCN TXT
These edge models can only hear; attempting to see will crash Ollama, but this can trim down on memory for hearing operations only. These edge models can only see; attempting to hear will crash Ollama, but this can trim down on memory for viewing operations only. If we only want the tool use and reasoning that Google’s Gemma4 edge models can provide with non-multimodal functionality (low priority) [basically fully stripped model]; attempting to use any multimodal functions would crash this so it is not recommended over the standard models.

Space saving with E2B: Sound only saves 6.82% of storage; while Vision only saves 13.64% of storage.

Space saving with E4B: Sound only saves 5.24% of storage; while Vision only saves 9.21% of storage.

This might seem counter intuitive to my focus for proper porting, but allows to save memory by only downloading and loading multi-modal functions what is needed for a project. If you don’t use the trimmed model beyond it’s focus, Ollama will never crash and you can benefit with loading tensors you actually need (also trims the file down a bit without harming quality). 👏

example: fredrezones55/Gemma-4-Uncensored-HauhauCS-Aggressive:e2b-SND; this model hearing functions fully work, tooling, and thinking. But vision was trimmed out for any projects that will never use it; in comparison fredrezones55/Gemma-4-Uncensored-HauhauCS-Aggressive:e2b-SCN this model has vision fully working but it’s hearing functions were trimmed for projects that will never use sound inputs.

Will expand to others later… Also waiting for HauhauCS to post the other sizes.