llmfan46/Qwythos-9B-Claude-Mythos-5-1M-uncensored-heretic

VISIONConcurrency Cost:1Model Size:9BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Jun 29, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

llmfan46/Qwythos-9B-Claude-Mythos-5-1M-uncensored-heretic is a 9 billion parameter language model, a decensored version of Empero AI's Qwythos-9B-Claude-Mythos-5-1M. It significantly reduces refusal rates (11/100 vs 73/100) while maintaining model quality, achieved using the Heretic v1.2.0 framework and Magnitude-Preserving Orthogonal Ablation. This model is optimized for reasoning tasks, offering a 1M-token context window and native function calling capabilities for enhanced factual accuracy and complex problem-solving.

Loading preview...

Qwythos-9B-Claude-Mythos-5-1M-uncensored-heretic Overview

This model is a decensored 9 billion parameter version of Empero AI's Qwythos-9B-Claude-Mythos-5-1M, created using the Heretic v1.2.0 framework and a variant of the Magnitude-Preserving Orthogonal Ablation (MPOA) method. It achieves a substantial reduction in refusal rates, dropping from 73/100 to 11/100, with minimal impact on model quality (0.0123 KL divergence).

Key Capabilities

  • Decensored Output: Significantly fewer refusals while preserving the original model's quality.
  • 1M-token Context Window: Features YaRN rope-scaling for a 1,048,576-token context, enabling whole-codebase reasoning, multi-document research, and long agentic trajectories.
  • Native Function Calling: Supports OpenAI/Qwen3.5-style function calling out-of-the-box, allowing self-correction with tools like Python executors and web search.
  • Enhanced Reasoning: Post-trained on over 500 million tokens of Claude Mythos and Fable traces, with in-house generated chain-of-thought, leading to strong performance in MMLU and GSM8K benchmarks.

Good For

  • Applications requiring uncensored responses for technically demanding questions in cybersecurity, red-teaming, biology, pharmacology, and clinical medicine.
  • Use cases benefiting from long context windows, such as analyzing large codebases or extensive research documents.
  • Agentic systems that leverage tool use and self-correction for improved factual accuracy and complex problem-solving.

Finetunes
2 models