Llmfan46 Frees Gemma-4-31B-it-qat-q4_0-unquantized-uncensored-heretic

    
        By vramkickedin    
     | 
    
            June 29, 2026 at 4:03 pm        
    
     | 
    
        2 min read

The new release called Gemma-4-31B-it-qat-q4_0-unquantized-uncensored-heretic is a modified version of the Gemma 4 language model designed to bypass safety filters. It uses quantization-aware training to maintain performance while reducing memory requirements. This specific release provides half-precision weights extracted from the training pipeline for custom downstream compilation and research.

Developer Llmfan46 who recently released the Gemma-4-Harmonia-31B-uncensored-heretic model created this uncensored version using a method called Abliteration to remove content restrictions. They host over 70 free models as an independent contributor but have reached the free storage limit on their platform. Without financial support to cover additional storage fees, no more new models can be uploaded.

Removing restrictions while maintaining logic

Key Features

Features half-precision weights for custom research.
Uses Abliteration method to remove content filters.
Reduces refusals by eighty-nine percent total.
Promises low divergence to preserve model quality.

People who need an AI assistant without built-in content restrictions will find this tool useful for open research and development. Anyone running models locally can use the provided files to test custom workflows without hitting standard safety roadblocks. It is also helpful for those who want to compile and modify the underlying weights for specific hardware setups.

Developer notes and project status

The developer uses targeted components to alter the model without completely destroying its original behavior. Testing shows the modified version drops accuracy slightly on the MMLU benchmark compared to the base model. The creator notes that doing all this took many days of work and effort for the community to make good use of these models.