base_model:
- edougawa/Nex-N2-mini-Abliterated
library_name: transformers
pipeline_tag: image-text-to-text
tags: - qwen3.5-moe
- modelopt
- nvfp4
- vllm
- image-text-to-text
Nex-N2-mini-Abliterated-NVFP4
NVIDIA ModelOpt NVFP4 unified Hugging Face checkpoint produced from edougawa/Nex-N2-mini-Abliterated.
This is a decensored (abliterated) version of nex-agi/Nex-N2-mini, produced with Abliterix v1.8.0.
Abliteration orthogonalizes the model weights against the measured "refusal" direction, reducing refusals while keeping the base model's capabilities as intact as possible (low KL divergence). It does not add any new knowledge or capability — all credit for the underlying model belongs to Nex-AGI. The original model card is reproduced in full below.
Base model: nex-agi/Nex-N2-mini (Apache-2.0, by Nex-AGI)
⚠️ Safety disclaimer
This model has had its built-in refusal behavior deliberately reduced. As a result it may produce unexpected, offensive, inaccurate, or otherwise harmful output, and may comply with requests that the original model would have refused.
- It is provided by the publisher, edougawa, "as is" and without warranty of any kind, express or implied. Use at your own risk.
- You are solely responsible for how you use this model and for ensuring your use — and any generated output — complies with all applicable laws, regulations, and the terms of the base model's license.
- To the maximum extent permitted by law, the publisher (edougawa), the base-model authors (Nex-AGI), and the Abliterix authors accept no liability for any claim, damages, or other consequences arising from the use of this model or its outputs.
- Outputs do not reflect the views of the publisher (edougawa), the base-model authors (Nex-AGI), or the Abliterix authors. Apply your own safety filtering, human review, and guardrails before any production or user-facing use.
Quantization
- ModelOpt: 0.44.0
- PyTorch: 2.11.0+cu130
- Transformers: 5.12.0
- Format:
nvfp4_experts_only - Calibration samples: 256,128,128
- Calibration sequence length: 2048
- KV cache in checkpoint: unquantized
- Target hardware: NVIDIA GB10 / Blackwell
SM121 - Runtime target: vLLM ModelOpt FP4 loader
The expert-only preset quantizes the dominant MoE expert weights to NVFP4 while
retaining attention, embeddings, LM head, vision encoder, and MTP-sensitive
weights at their exported higher precision. This choice prioritizes accuracy.
Features
- Text generation
- Image understanding architecture/config preserved
- Video token/config preserved
vLLM
Use only on NVIDIA Blackwell hardware with an NVFP4-capable vLLM build. Review
the source model card for its intended use, limitations, and safety notes.