LLMFan46's uncensored Gemma-4 finetune offers low KLD and few refusals
New 26B A4B model runs faster on smaller VRAM with only 12% refusal rate.
Deep Dive
LLMFan46 has released the G4-MeroMero-26B-A4B-it-uncensored-heretic, a smaller, faster uncensored finetune for users with limited VRAM or RAM, available in Safetensors and GGUF formats on Hugging Face.
Key Points
- KLD of 0.0152 ensures high similarity to base Gemma-4-26B-A4B-it while reducing censorship
- Only 12 refusals per 100 prompts, making it one of the least censored finetunes
- Available as both Safetensors and GGUF for use with various inference engines
Why It Matters
Gives professionals access to an uncensored AI model that runs efficiently on modest hardware.