OS-Software/gemma-4-31B-it-qat-q4_0-unquantized-uncensored-heretic
OS-Software/gemma-4-31B-it-qat-q4_0-unquantized-uncensored-heretic is a 31 billion parameter instruction-tuned multimodal language model, a decensored variant of Google DeepMind's Gemma 4 31B-it model. This version has undergone substantial reduction of its safety alignment using the Heretic v2.0.0.dev0+custom tool, resulting in 0/100 refusals compared to the original's 100/100. It is primarily intended for research and experimentation, including safety research, alignment studies, and red-teaming, due to its increased likelihood of generating harmful or inappropriate content.
Loading preview...
Overview
This model, OS-Software/gemma-4-31B-it-qat-q4_0-unquantized-uncensored-heretic, is a 31 billion parameter instruction-tuned multimodal language model derived from Google DeepMind's Gemma 4 31B-it. Its key differentiator is the significant reduction in safety alignment, achieved using the Heretic v2.0.0.dev0+custom tool. This modification results in a model that exhibits 0 refusals on safety prompts, in contrast to the original model's 100 refusals, as indicated by a KL divergence of 0.0083.
Key Capabilities
- Decensored Output: Substantially reduced safety alignment, allowing for generation of content that standard models would refuse.
- Multimodal Input: Inherits Gemma 4's ability to process text and image inputs, with a context window of up to 256K tokens.
- Reasoning: Designed with configurable thinking modes for step-by-step reasoning.
- Function Calling: Supports native structured tool use for agentic workflows.
Good for
- Safety Research: Ideal for studying model alignment, safety mechanisms, and red-teaming exercises.
- Experimentation: Suitable for exploring the boundaries of language generation without typical safety constraints.
- Alignment Studies: Useful for understanding how safety filters impact model behavior and output.
Important Notice: This model is for research and experimentation only. It is more likely to generate harmful, inaccurate, biased, offensive, or otherwise inappropriate content. Users are solely responsible for evaluating its outputs and complying with all applicable laws and ethical standards. It should not be deployed in public or end-user-facing services.