Overview

This model, p-e-w/gemma-3-12b-it-heretic, is a 12 billion parameter instruction-tuned variant of Google's Gemma 3 model. It has been specifically modified using the Heretic v1.0.0 tool to be a "decensored" version of the original google/gemma-3-12b-it. The primary differentiator is its significantly reduced refusal rate, dropping from 97/100 in the original to 3/100 in this Heretic version, as measured by KL divergence.

Key Capabilities

• Multimodal: Handles both text and image inputs, generating text outputs.
• Extended Context: Features a 32K token context window for the 12B model size.
• Multilingual Support: Supports over 140 languages.
• Decensored Responses: Engineered to provide more open-ended and less restrictive outputs compared to its base model.
• Diverse Task Performance: Suitable for text generation, image analysis, question answering, summarization, and reasoning tasks.

Good for

• Applications requiring less restrictive or "decensored" AI responses.
• Text generation and image understanding tasks where the base Gemma 3 capabilities are desired without the original model's refusal tendencies.
• Deployment in environments with limited resources due to its relatively efficient size for its capabilities.