DavidAU/gemma-4-31B-it-Mystery-Fine-Tune-HERETIC-UNCENSORED-Thinking

VISIONConcurrency Cost:2Model Size:31BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Apr 3, 2026License:apache-2.0Architecture:Transformer0.1K Open Weights Cold

DavidAU's gemma-4-31B-it-Mystery-Fine-Tune-HERETIC-UNCENSORED-Thinking is a 31 billion parameter Gemma 4 instruction-tuned model, fine-tuned by DavidAU using private datasets with Unsloth. This model is designed for 'thinking mode' applications, demonstrating strong performance in benchmarks, exceeding the root model in 6 out of 7 tests, and is characterized by its 'heretic/uncensored' nature. It is optimized for complex reasoning tasks and creative generation, particularly for scenarios requiring unfiltered and detailed responses.

Loading preview...

Model Overview

This is a Gemma 4 31B instruction-tuned model, developed by DavidAU, utilizing a private dataset and the Unsloth fine-tuning framework. It is explicitly labeled as "HERETIC-UNCENSORED" and is designed with a strong emphasis on a "thinking mode" for generating detailed, step-by-step reasoning. The model demonstrates robust performance, outperforming the base Gemma 4 31B model in 6 out of 7 internal benchmarks, including arc-c, arc/e, hswag, obkqa, piqa, and wino.

Key Capabilities

  • Enhanced Reasoning: Features a built-in "thinking mode" (<|channel>thought...<channel|>) for step-by-step problem-solving and detailed internal monologue.
  • Uncensored Output: Fine-tuned to provide unfiltered responses, suitable for use cases requiring less restrictive content generation.
  • Strong Benchmark Performance: Exceeds the root Gemma 4 31B model in most evaluated benchmarks, indicating improved instruction following and reasoning.
  • Long Context: Inherits the Gemma 4's 128K token context window, enabling processing of extensive inputs.
  • Multimodal Foundation: Based on Gemma 4, which supports advanced capabilities like image understanding, video understanding, and interleaved multimodal input (though this specific fine-tune focuses on text).

Good For

  • Applications requiring detailed, multi-step reasoning and internal thought processes.
  • Creative writing and role-playing scenarios where uncensored and 'heretic' responses are desired.
  • Developers seeking a high-performing Gemma 4 variant for instruction-tuned tasks.
  • Use cases that benefit from a large context window for complex prompts.

Finetunes
2 models