Gemma-4-Harmonia-31B-Uncensored-Heretic Is Out Now, a Merge of Multiple gemma-4-31B-it Finetunes Designed for a Targeted Approach to Deep Neural Consolidation, Minimizing Regression While Amplifying Unique Capability Boundaries. With KLD 0.0047 and 9/100 Refusals!

Posted by LLMFan46@reddit | LocalLLaMA | View on Reddit | 9 comments

Provided in both Safetensors and GGUFs.

Safetensors, llmfan46/Gemma-4-Harmonia-31B-it-uncensored-heretic: https://huggingface.co/llmfan46/Gemma-4-Harmonia-31B-it-uncensored-heretic

GGUFs, llmfan46/Gemma-4-Harmonia-31B-it-uncensored-heretic-GGUF: https://huggingface.co/llmfan46/Gemma-4-Harmonia-31B-it-uncensored-heretic-GGUF

Comes with benchmark too.

Find all my models here: HuggingFace-LLMFan46

The original author of this finetune is: virtuous7373

[-]

ExoticYesterday8282@reddit

Minimizing regression while amplifying unique capability boundaries
bro summoned the most academic way possible to say
we made Gemma less scared. 😭

Eat-Playdoh@reddit

Clanker hands typed that 🤖

Glittering_Focus1538@reddit

Word filler fr

New_Comfortable7240@reddit

OP one question, is this finetune focused on character cards, on short stories, on ST? Or what can you say it's their strengths?

tarruda@reddit

This seems to be just an experiment, even the original uploader didn't test it much: https://huggingface.co/virtuous7373/Gemma-4-Harmonia-31B/discussions/1#6a10bc5811f02f3df7f05ae8

Thanks for sharing! Will give it a try

Spara-Extreme@reddit

What is this finetune actually providing over just heretic ?

looselyhuman@reddit

Has anyone worked with the Harmonia model? Seems really interesting? I get the feeling it could be a little neurotic from all the creative models that went into it, but probably fun to talk to.

trackinstats@reddit

Whoa