Open Source

Gemma-4-Harmonia-31B: Uncensored merge minimizes regression, amplifies capabilities

KLD 0.0047 and only 9 refusals per 100 prompts — a new open-weights finetune arrives

Deep Dive

LLMFan46 shares a new finetune by virtuous7373: Gemma-4-Harmonia-31B-uncensored-heretic. Available in both Safetensors and GGUF formats on HuggingFace, with benchmarks included. Find all models on HuggingFace-LLMFan46.

Key Points
  • Merge of multiple Gemma-4-31B-it finetunes with KLD 0.0047 to minimize regression
  • Uncensored variant with only 9 refusals per 100 prompts
  • Available in Safetensors and GGUF formats on HuggingFace from LLMFan46

Why It Matters

A specialized, low-refusal model for developers needing uncensored creativity without sacrificing performance.