Qwen3.5-9B-abliterated

A fully uncensored version of Qwen3.5-9B using orthogonal projection abliteration.

Playable

Description

This model is an uncensored variant of Qwen3.5-9B that removes refusal behaviors through a two-stage approach of orthogonal projection abliteration and LoRA fine-tuning. It achieves a 100% success rate on refusal benchmarks while retaining reasoning, coding, and general knowledge capabilities.

Descriptions, tags, and model credits may be AI-generated or inferred from public sources and can be incomplete or wrong. Learn more.

Details

Why this is interesting: Demonstrates a successful two-stage 'abliteration' technique to achieve 100% uncensored behavior on a 9B model while…
Attributed to: jeric caballegan@lukey03
Built with: Nano Banana 2 Lite Qwen3.5 9B Qwen3.5 4B Qwen3.5 0.8B Seedance 2.0 Qwen3 14B Qwen3 30B-A3B Qwen3 8B w2v-BERT 2.0
Links: huggingface.co
View tweet
Tags: #uncensored #abliterated #llm #fine-tuned

Discovered: X
First posted: X · Sep 25, 2026
Listed: Added 7d ago
Build: Strong
Models: Strong