DreamFast/gemma-3-12b-it-heretic

DreamFast/gemma-3-12b-it-heretic: Decensored Gemma 3 12B IT

This model is an 'abliterated' version of Google's Gemma 3 12B IT, processed using the Heretic v1.1.0 tool. Its primary distinction is a significant reduction in refusal rates, dropping from 100/100 in the original base model to just 7/100, indicating a 93% success rate for previously refused prompts. This decensoring process aims to remove 'soft censorship' that can weaken or sanitize concepts in embeddings.

Key Characteristics & Optimizations

• Reduced Refusals: Achieves a refusal rate of 7/100, compared to 100/100 for the base Gemma 3 12B IT model.
• Minimal Model Damage: The abliteration process resulted in a low KL Divergence of 0.0826, suggesting that the core model quality is largely preserved.
• Targeted for Video Generation: Specifically designed as an uncensored text encoder for video generation models like LTX-2. By removing inherent censorship, it enables more faithful adherence to creative prompts and prevents softened or altered visual outputs.
• Multiple Formats: Available in HuggingFace, ComfyUI (bf16, fp8), and various GGUF quantizations (F16, Q8_0, Q6_K, Q5_K_M, Q4_K_M recommended).

Limitations

• Inherits all limitations of the base Gemma 3 12B model.
• This v1 model lacks vision capabilities; v2 is available for vision support.
• Abliteration reduces, but does not entirely eliminate, refusals.

This model is ideal for developers requiring a less restrictive text encoder for creative applications, particularly in video generation workflows where prompt fidelity is crucial.