coder3101/gemma-4-31B-it-heretic

Overview

This model, coder3101/gemma-4-31B-it-heretic, is a 31 billion parameter instruction-tuned variant of Google's Gemma 4 family, specifically a decensored version of google/gemma-4-31B-it. It was created using the Heretic tool with the Arbitrary-Rank Ablation (ARA) method, aiming to reduce content refusals while preserving core capabilities. The model supports a substantial context window of up to 256K tokens (though this specific variant is noted with 32768 tokens in the prompt) and handles both text and image inputs.

Key Differentiators

• Decensored Behavior: Achieves a refusal rate of 15/100 compared to the original model's 99/100, indicating a significantly less restrictive output. This is a direct result of the Heretic abliteration process.
• Multimodal Capabilities: Inherits the Gemma 4 family's ability to process text and image inputs, with support for variable aspect ratio and resolution images.
• Reasoning and Coding: Designed for strong reasoning, agentic workflows, and enhanced coding capabilities, including native function-calling support.
• Extended Context: Features a large context window, enabling processing of complex, long-context tasks.

Should I use this for my use case?

This model is particularly suitable if your application requires a powerful, multimodal language model with a significantly reduced tendency for content refusals, especially compared to its original Google counterpart. It excels in tasks requiring reasoning, code generation, and multimodal understanding (text and image). Developers needing a less restrictive model for creative or open-ended applications, while still benefiting from the Gemma 4 architecture, would find this model highly relevant. However, users should be aware of the ethical implications and potential risks associated with a decensored model and implement appropriate safety safeguards.