llmfan46/Tower-Plus-72B-ultra-uncensored-heretic

Hugging Face
TEXT GENERATIONConcurrency Cost:4Model Size:72.7BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Jun 14, 2026License:cc-by-nc-sa-4.0Architecture:Transformer Open Weights Warm

llmfan46/Tower-Plus-72B-ultra-uncensored-heretic is a 72.7 billion parameter language model developed by llmfan46, based on Unbabel's Tower-Plus-72B. This model has been decensored using the Heretic v1.4.0 framework and a Magnitude-Preserving Orthogonal Ablation (MPOA) method, significantly reducing refusals while maintaining original model quality. It is primarily designed for multilingual tasks, excelling in translation-related tasks across 22 languages, and is built on Qwen 2.5 72B.

Loading preview...

Model Overview

llmfan46/Tower-Plus-72B-ultra-uncensored-heretic is a 72.7 billion parameter model derived from Unbabel's Tower-Plus-72B, which is built upon Qwen 2.5 72B. This version has been specifically modified by llmfan46 using the Heretic v1.4.0 framework and a Magnitude-Preserving Orthogonal Ablation (MPOA) method to significantly reduce content refusals.

Key Capabilities & Differentiators

  • Decensored Performance: Achieves 95% fewer refusals (5/100) compared to the original model (100/100) while preserving model quality with a low KL divergence of 0.0516.
  • Multilingual Expertise: The base Tower+ model is fine-tuned on a mix of translation-related tasks and general instruction-following datasets, covering 22 languages including German, Spanish, French, Italian, Korean, Dutch, Russian, English, Portuguese, Chinese, and Japanese.
  • Context Length: Supports a substantial context size of 131,072 tokens, with a recommended generation token limit of 8192.

Intended Use Cases

  • Multilingual Tasks: Particularly strong in translation-related tasks across its supported languages.
  • Synthetic Data Generation: Effective for creating multilingual synthetic data, either by translating instructions and answers or generating instructions from seed documents.
  • Uncensored Applications: Suitable for use cases requiring a model with significantly reduced content restrictions and refusals, without substantial degradation in baseline performance.

Finetunes
1 models