llmfan46/Nex-N2-mini-ultra-uncensored-heretic

TEXT GENERATIONConcurrency Cost:3Model Size:35.1BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Jun 24, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

The llmfan46/Nex-N2-mini-ultra-uncensored-heretic is a 35.1 billion parameter language model developed by llmfan46, based on the nex-agi/Nex-N2-mini architecture. This model is a decensored variant created using Heretic v1.2.0 with the Magnitude-Preserving Orthogonal Ablation (MPOA) method, significantly reducing refusals by 93% while maintaining original model quality with a 0.0020 KL divergence. It is optimized for agentic workflows, coding tasks, and general reasoning, offering enhanced utility for applications requiring fewer content restrictions.

Loading preview...

Nex-N2-mini-ultra-uncensored-heretic Overview

This model, developed by llmfan46, is a 35.1 billion parameter decensored version of the nex-agi/Nex-N2-mini. It was created using the Heretic v1.2.0 framework and the Magnitude-Preserving Orthogonal Ablation (MPOA) method to specifically address content refusal rates.

Key Differentiators & Performance

  • Decensored Output: Achieves a significant 93% reduction in refusals (5/100 compared to 74/100 for the original model), making it suitable for use cases requiring less content restriction.
  • Quality Preservation: Maintains high fidelity to the original Nex-N2-mini, evidenced by a low KL divergence of 0.0020.
  • Agentic Capabilities: The base Nex-N2 architecture is designed for "Agentic Thinking," unifying reasoning, tool use, and environment execution through Adaptive and Coherent Thinking frameworks.
  • Coding & General Reasoning: Excels in agentic workflows, coding tasks, and general reasoning, with the original Nex-N2-mini showing strong performance across benchmarks like BrowseComp (74.1), Terminal-Bench 2.1 (60.7), and GPQA Diamond (82.6).
  • Function Calling: Supports robust function-calling capabilities, configurable during server launch.

Targeted Components for Decensoring

The MPOA method specifically targeted and modified the attn.o_proj, attn.out_proj, and mlp.down_proj components to achieve the reduction in refusals.

Usage Recommendations

For optimal performance, it is recommended to use the customized sglang fork provided by Nex-AGI for deployment. Suggested sampling parameters include temperature: 0.7, top_p: 0.95, and top_k: 40.

Finetunes
2 models