MuXodious/gpt-oss-20b-RichardErkhov-heresy

TEXT GENERATIONConcurrency Cost:1Model Size:20BQuant:FP8Ctx Length:32kPublished:Feb 7, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

MuXodious/gpt-oss-20b-RichardErkhov-heresy is a 20 billion parameter gpt-oss-20b fine-tune, created by MuXodious using a modified Heretic ablation engine with Magnitude-Preserving Orthogonal Ablation. This model is specifically engineered for extreme decensorship, achieving a low refusal rate while minimizing damage to the original model's capabilities. It excels in scenarios requiring highly unconstrained and direct responses, topping the UGI leaderboard for models under 24B in willingness scores.

Loading preview...

Model Overview

MuXodious/gpt-oss-20b-RichardErkhov-heresy is a 20 billion parameter model derived from OpenAI's gpt-oss-20b base, fine-tuned by MuXodious. It was developed using a highly modified version of P-E-W's Heretic ablation engine, specifically incorporating Magnitude-Preserving Orthogonal Ablation. This process aims to significantly reduce content refusal rates while preserving the model's core functionality.

Key Characteristics

  • Extreme Decensorship: Engineered to be highly unconstrained, achieving a refusal rate of 6/100, a substantial reduction from an initial 98/100 refusals.
  • Performance: As of May 2026, it leads the UGI leaderboard for models under 24B in willingness scores.
  • Base Model Capabilities: Inherits the gpt-oss-20b features, including configurable reasoning effort (low, medium, high), full chain-of-thought access, and agentic capabilities like function calling, web browsing, and Python code execution.
  • Quantization: Utilizes MXFP4 quantization, allowing the 20B parameter model to run within 16GB of memory.
  • Permissive Licensing: Based on a model with an Apache 2.0 license, suitable for experimentation, customization, and commercial deployment.

Ideal Use Cases

  • Unfiltered Content Generation: Suited for applications requiring highly direct and uncensored responses.
  • Research into Model Alignment/Safety: Useful for studying the effects of ablation and decensorship techniques on LLMs.
  • Agentic Workflows: Can be leveraged for tasks involving tool use, web browsing, and code execution due to its gpt-oss lineage.
  • Specialized Fine-tuning: The gpt-oss-20b base is fine-tunable on consumer hardware for specific use cases.

Finetunes
3 models