Gemma-4-31B-it-Mystery-Fine-Tune-HERETIC-UNCENSORED-Thinking Now Live

By vramkickedin | | 2 min read
Angled geometric glass structure revealing three distinct translucent panels stacked and fanning outward.

A newly released open weight model, Gemma-4-31B-it-Mystery-Fine-Tune-HERETIC-UNCENSORED-Thinking, removes standard content filters from a 31B parameter architecture while activating a built-in reasoning workflow. This configuration produces step-by-step internal logic before generating final text outputs.

Developer DavidAU trained the initial version on private datasets using specialized optimization tools to establish a testing baseline. The project provides a transparent framework for local inference without default safety interventions.

Model Size: 60GB+ & VRAM GPU: requirements vary

Core functions and workflow

  • Forces continuous reasoning output between designated channel tags before delivering final answers.
  • Bypasses standard content restrictions across creative and technical prompts.
  • Maintains a 128,000-token context window inherited from the original architecture.
  • Outperforms the untouched base model in six out of seven standard evaluation tests.

Writers generating unfiltered fiction, researchers prototyping unrestricted dialogue systems, and teams testing local pipelines will find this configuration useful for evaluating raw model behavior.

Training adjustments and version roadmap

The initial release serves as a test case for handling newer architectural patterns that combine elements from several recent designs. Subsequent iterations will apply lighter tuning weights to preserve base capabilities while refining the reasoning output structure. Manual reviews confirmed the removal of standard filters, though minor formatting inconsistencies around the thought tags may require template adjustments in third-party software. V2 training begins shortly after the current upload, with benchmark data expected once the new weights become available.

DavidAU observed that

"V1 came out very strong ; it maybe perfect for your use cases or too strong,"

Future builds will adjust the intensity based on how well the reasoning blocks align with standard expectations.

This configuration provides a transparent look at unrestricted reasoning behavior for local deployment workflows. Download the Gemma-4-31B-it-Mystery-Fine-Tune-HERETIC-UNCENSORED-Thinking weights here.