spinochenza/Nex-N2-mini-ultra-uncensored-heretic

TEXT GENERATIONConcurrency Cost:3Model Size:35.1BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Jul 2, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

The spinochenza/Nex-N2-mini-ultra-uncensored-heretic is a 35.1 billion parameter language model based on the nex-agi/Nex-N2-mini architecture, which itself is built on the Qwen3.5 series. This model has been decensored using the Heretic v1.2.0 framework with a Magnitude-Preserving Orthogonal Ablation (MPOA) method, resulting in 93% fewer refusals (5/100) compared to the original while maintaining a low KL divergence of 0.0020. It is specifically optimized for reduced content restrictions, making it suitable for applications requiring less filtered responses.

Loading preview...

Model Overview

This model, spinochenza/Nex-N2-mini-ultra-uncensored-heretic, is a 35.1 billion parameter variant of the nex-agi/Nex-N2-mini model, which is part of the Nex-N2 family built on the Qwen3.5 series. Its primary distinction is its decensored nature, achieved through the application of the Heretic v1.2.0 framework using a variant of the Magnitude-Preserving Orthogonal Ablation (MPOA) method.

Key Differentiators

  • Significantly Reduced Refusals: Achieves a remarkable 93% reduction in refusals, with only 5 out of 100 test cases resulting in a refusal, compared to 74 out of 100 for the original Nex-N2-mini. This indicates a much lower propensity for content restrictions, objections, or censorship.
  • Preserved Quality: Despite the decensoring process, the model maintains high fidelity to the original's baseline performance, evidenced by a very low KL divergence of 0.0020.
  • Targeted Ablation: The decensoring process specifically targeted components like attn.o_proj, attn.out_proj, and mlp.down_proj to achieve its uncensored behavior.

Ideal Use Cases

  • Applications requiring less filtered or unrestricted text generation.
  • Scenarios where the original Nex-N2-mini's refusal rate was a limiting factor.
  • Research into model safety, bias, and control mechanisms.

For local deployment, nex-agi recommends using their customized sglang fork, with specific configurations for Nex-N2-mini including qwen3 reasoning and qwen3_coder tool-call parsers. GGUF quantizations are also available for broader compatibility.

Finetunes
1 models