saidutta69/gemma-3-12b-it-heretic
The saidutta69/gemma-3-12b-it-heretic is a 12.2 billion parameter decensored variant of Google's Gemma 3 12B instruction-tuned model, featuring a 32768 token context length. Developed by RACER IS OP using the Heretic v1.4.0 'abliteration' method, it suppresses refusal behavior via targeted weight edits rather than fine-tuning, preserving the base model's knowledge and instruction-following. This multimodal model is designed for developers seeking an uncensored Gemma 3 12B for local agents, image-grounded Q&A, and research into alignment and refusal mechanics.
Loading preview...
Overview
saidutta69/gemma-3-12b-it-heretic is a 12.2 billion parameter multimodal language model derived from Google's gemma-3-12b-it. Its primary distinction is the removal of refusal guardrails through a technique called abliteration (Heretic v1.4.0). This method directly edits specific weight directions responsible for refusal in the attention output and MLP down-projections, aiming to leave the base model's core knowledge and instruction-following capabilities largely intact, unlike traditional fine-tuning which can degrade coherence.
Key Capabilities and Features
- Decensored Behavior: Significantly reduced refusal rates (from 99/100 to 40/100 on adversarial prompts) compared to the base Gemma 3 12B model.
- Multimodal: Inherits the text and vision capabilities of the original Gemma 3 12B.
- Weight Editing: Utilizes directional ablation to suppress refusal, preserving the base model's original training.
- GPU Compatibility: Optimized GGUF quantizations (Q8_0, Q6_K, Q5_K_M, Q4_K_M, IQ4_XS) are provided to run on GPUs with VRAM as low as 6GB, including consumer-grade gaming PCs.
- Context Length: Supports the base model's 32768 token context window.
Use Cases
This model is particularly suited for:
- Developers requiring an uncensored, capable multimodal model for local agents.
- Applications involving image-grounded Q&A without content restrictions.
- Research into model alignment, refusal mechanisms, and the effects of targeted weight editing.
Note: As a decensored model, it will comply with requests the base model would refuse. Users are responsible for its deployment and should not expose it via unmoderated public-facing endpoints.