llmfan46/Magistral-Small-2509-ultra-uncensored-heretic-v2

VISIONConcurrency Cost:2Model Size:24BQuant:FP8Ctx Length:32kPublished:Mar 17, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

The llmfan46/Magistral-Small-2509-ultra-uncensored-heretic-v2 is a decensored version of Mistral AI's 24B parameter Magistral-Small-2509 model, created using the Heretic v1.2.0 tool with Arbitrary-Rank Ablation. This model significantly reduces refusals (3/100 compared to 92/100 in the original) while preserving core model quality with a low KL divergence of 0.0493. It retains the original model's advanced reasoning, multilingual, and vision capabilities, making it suitable for applications requiring less content restriction.

Loading preview...

Overview

This model, llmfan46/Magistral-Small-2509-ultra-uncensored-heretic-v2, is a decensored variant of Mistral AI's 24-billion parameter Magistral-Small-2509. It was developed using the Heretic v1.2.0 tool, specifically employing the Arbitrary-Rank Ablation (ARA) method to modify its behavior. The primary goal of this modification is to drastically reduce content refusals while maintaining the original model's performance.

Key Differentiators

  • Significantly Reduced Refusals: Achieves a refusal rate of only 3/100, a substantial improvement over the original model's 92/100, making it highly "uncensored."
  • Preserved Quality: Despite decensoring, the model maintains a low KL divergence of 0.0493, indicating minimal degradation of the original model's capabilities, coherence, and reasoning ability.
  • Enhanced Reasoning: Inherits the Magistral-Small-2509's advanced reasoning capabilities, including the ability to show its thinking process using [THINK]...[/THINK] tags.
  • Multimodal and Multilingual: Supports vision inputs, allowing it to analyze images and reason based on visual content, and is proficient in dozens of languages.
  • Large Context Window: Features a 128k context window, recommended for optimal performance up to 40k tokens.

Ideal Use Cases

  • Applications requiring a highly permissive and uncensored language model.
  • Tasks demanding complex reasoning and problem-solving, especially where the thinking process is beneficial.
  • Multilingual content generation and understanding.
  • Multimodal applications that involve both text and image analysis.

Finetunes
2 models