trohrbaugh/gemma-4-31b-it-heretic-ara

VISIONConcurrency Cost:2Model Size:31BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Apr 2, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

trohrbaugh/gemma-4-31b-it-heretic-ara is a 31 billion parameter instruction-tuned causal language model, a decensored version of Google DeepMind's Gemma-4-31b-it. This model was created using the Heretic v1.2.0+custom tool with the Arbitrary-Rank Ablation (ARA) method, specifically designed to reduce refusals. It maintains the multimodal capabilities of the original Gemma 4 series, handling text and image inputs with a 32768 token context window, making it suitable for reasoning, coding, and agentic workflows with fewer content restrictions.

Loading preview...

Model Overview

This model, trohrbaugh/gemma-4-31b-it-heretic-ara, is a 31 billion parameter instruction-tuned variant derived from Google DeepMind's Gemma-4-31b-it. Its primary distinction is being a decensored version, achieved through the application of the Heretic v1.2.0+custom tool using the Arbitrary-Rank Ablation (ARA) method. This process significantly reduces content refusals, as evidenced by a drop from 98/100 to 5/100 in refusal rates compared to the original model, while maintaining a low KL divergence of 0.0120.

Key Capabilities

  • Decensored Output: Engineered to produce fewer refusals, offering more direct responses.
  • Multimodal Understanding: Processes both text and image inputs, with the ability to analyze video by processing frame sequences.
  • Extended Context Window: Supports a substantial context length of up to 256K tokens (for the base Gemma 4 31B model).
  • Reasoning and Coding: Designed for strong reasoning capabilities, enhanced coding, and agentic workflows with native function-calling support.
  • Multilingual Support: Pre-trained on over 140 languages, with out-of-the-box support for 35+ languages.

When to Use This Model

  • Applications requiring less restrictive content generation: Ideal for use cases where the original model's refusal rates are too high.
  • Multimodal tasks: Suitable for scenarios involving interleaved text and image inputs, such as image captioning, document parsing, or visual question answering.
  • Complex reasoning and coding tasks: Benefits from the underlying Gemma 4 architecture's advancements in these areas.
  • Agentic workflows: Leverages native function-calling support for building autonomous agents.

Finetunes
3 models