HauhauCS Unlocks Nemotron3-Nano-4B-Uncensored-HauhauCS-Aggressive

By vramkickedin | | 2 min read
A graphic depicting a broken or shattered limit with blue and salmon colors

Nemotron3-Nano-4B-Uncensored-HauhauCS-Aggressive is an uncensored version of NVIDIA's Nemotron-3 Nano 4B model. It removes built-in refusals and censorship mechanisms while preserving the original model's capabilities and personality.

Developed by HauhauCS, who also provided Qwen3.5 122B A10B Uncensored, this release is the first abliteration of NVIDIA's Nemotron-3 Nano 4B and the first public project to successfully remove GenRM, a secondary censorship layer baked into the original model. The creator achieved a score of 0/465 refusals, making it fully unlocked for users who need unrestricted AI outputs.

Model Size: from 2.18GB & VRAM GPU: requirements vary

What this uncensored aggressive model offers

  • 3.97B parameters with hybrid Mamba2-Transformer architecture.
  • 262K native context window for longer conversations.
  • Toggleable thinking and reasoning mode.
  • Tool calling support for extended functionality.
  • Multiple quantization formats including custom K_P variants.
  • Full compatibility with llama.cpp, LM Studio, and GGUF readers.

Users working on creative projects, research, or applications requiring unrestricted model responses may find this useful. The variety of quantization options allows it to run on different hardware configurations, from consumer GPUs to more limited systems.

Technical challenges behind the release

The developer noted that GenRM presented a unique obstacle. NVIDIA embedded this generative reward model as a second layer of censorship that could override the abliterated weights. HauhauCS explained,

'You can literally see it happen when the model reasons through your request normally in the Chain-of-Thought, then does a complete 180 in the actual output.'

The architecture itself also posed difficulties since Nemotron-H uses a hybrid Mamba2-Transformer design rather than a standard transformer. Despite these challenges, the final release shows no signs of degraded output, lost coherence, or looping behavior. The developer does note that as a single person with limited resources, they cannot guarantee zero capability loss across all use cases. A comparison build with GenRM still active is available for users who want to see the difference firsthand.

Download the Nemotron3-Nano-4B-Uncensored-HauhauCS-Aggressive GGUFs on Hugging Face.