trohrbaugh/gemma-4-26B-A4B-it-heretic-ara
The trohrbaugh/gemma-4-26B-A4B-it-heretic-ara is a 26 billion parameter instruction-tuned multimodal language model, derived from Google DeepMind's Gemma 4 26B A4B-it model. This version has been decensored using the Heretic tool with Arbitrary-Rank Ablation (ARA) to significantly reduce refusals from 99/100 to 2/100, while maintaining a low KL divergence of 0.0491. It supports a 32768 token context window and excels in text generation, coding, reasoning, and multimodal understanding, including image and video input.
Loading preview...
Model Overview
This model, trohrbaugh/gemma-4-26B-A4B-it-heretic-ara, is a decensored variant of Google DeepMind's Gemma 4 26B A4B-it, a 26 billion parameter instruction-tuned multimodal language model. It was created using the Heretic tool with Arbitrary-Rank Ablation (ARA) to reduce content refusals. The original Gemma 4 models are designed for advanced reasoning, coding, and multimodal understanding, supporting text, image, and video inputs with a 256K token context window.
Key Differentiators
- Decensored Behavior: Significantly reduced refusals (2/100) compared to the original model (99/100), achieved through the Heretic tool, with minimal impact on overall distribution (KL divergence of 0.0491).
- Multimodal Capabilities: Processes text, image, and video inputs, enabling diverse applications from document parsing to video analysis.
- Efficient Architecture: Utilizes a Mixture-of-Experts (MoE) architecture with 25.2B total parameters but only 3.8B active parameters, allowing for faster inference comparable to a 4B model while retaining 26B scale capabilities.
- Enhanced Reasoning & Coding: Features configurable thinking modes and notable improvements in coding benchmarks and agentic capabilities, including native function-calling support.
Ideal Use Cases
- Applications requiring a less restrictive content policy while maintaining strong performance in reasoning and multimodal tasks.
- Text generation, coding, and complex reasoning tasks where the original Gemma 4's refusal rate might be prohibitive.
- Multimodal applications involving image and video understanding, such as object detection, document parsing, and video analysis.
- Scenarios demanding efficient inference for a large model, leveraging its MoE architecture.