abliteration

This article presents a comprehensive benchmark of 8 abliterated variants of the Qwen 3.8 27B model against the base model, using weight analysis, KL divergence, 13 benchmarks, and HarmBench refusal tests over 167 GPU hours. The analysis reveals that surgical edits significantly outperform heavy modifications, with aggressive abliteration causing thinking loops in up to 45% of adversarial responses and chat template manipulation detected in some variants.

Abliteration.ai is a startup that commercializes the removal of AI model guardrails, offering unfiltered access to models like GLM-5.3 for offensive cyber and red-teaming work, while sparking debates on AI safety and potential misuse.

This article describes an uncensored version of Z.ai's GLM-5.3-Flash model, with safety alignments removed via abliteration, released as a block-FP8 checkpoint for research purposes.

The Qwen3.8-27B-Uncensored is a 27B-parameter AI model with safety alignment removed via abliteration, released on Hugging Face for research purposes without built-in guardrails.

Ornith 1.5 9B Abliterated, an experimental MLX derivative, is now available in 4-bit, 8-bit, and BF16 builds for Apple Silicon, designed to reduce refusal behavior while maintaining capabilities for research and legitimate local use.

Alibaba's Qwen3.8-27B model's safety restrictions have been completely removed, with the V2 version achieving zero rejection, and the MMLU score improved to 86.3%, using complementary ablation blending to achieve uncensored and stronger performance.

A new uncensored AI model, Qwen3.8-27B-OBLITERATED, is released with zero refusal on harmful prompts, optimized for cybersecurity tasks and jailbreaking, featuring a novel abliteration blending technique.

The paper introduces AMRA, a weight-editing method to mitigate abliteration in large language models by obscuring the refusal signal, improving post-abliteration refusal scores on Llama-3-8B and Gemma-2-9B with minimal utility degradation.

OBLITERATUS has released a modified version of Alibaba's Qwen3.8-27B model with all safety refusals removed, achieving zero refusals across 842 harmful prompts through iterative surgical modifications.

This is an abliterated (refusal-removed) version of the Qwen3.8-27B AI model, released for research purposes like interpretability and red-teaming, with warnings about its lack of safety guardrails.

This article presents an uncensored version of the Qwen3.8 27B AI model, which has been modified to remove safety refusals and quantized to FP8 for efficiency, intended for research purposes.

An uncensored variant of the Qwen3.8-27B LLM, modified via abliteration to remove content restrictions, provided in GGUF format for deployment with llama.cpp and ollama.

This is an uncensored version of the Qwen3.8-27B AI model created using abliteration to remove refusals, serving as a proof-of-concept for modifying LLMs without extensive tools.

This is a GGUF conversion of the uncensored Qwen3.8-27B model, an abliterated version with safety alignment removed for research purposes, running on llama.cpp with various quantizations and vision support.

This is an early access draft of an uncensored, abliterated version of the Qwen3.8-27B AI model, designed to remove safety censorship while maintaining coherence, though it has edge cases with long-context generation.

This article describes the release of Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF, a double-refined abliterated variant of the Qwen model with reduced refusals for adult audiences, using ARA technique for research and creative writing.

Scotoma-2 is an updated fine-tune of Gemma-4-31B-it that reduces repetitive writing tics via targeted preference training while keeping the base model's intelligence. It is not uncensored but aims to produce cleaner, less annoying prose for roleplay.

A study on uncensored LLMs (Gemma and Qwen) shows that removing censorship makes them more optimistic in stock market predictions, but not more accurate. The effect varies by model family.

A comparison of 23 Gemma 4 E4B models on HuggingFace shows that the most downloaded model, OBLITERATUS, is completely broken, while the more surgical 'heretic' variants perform best.

Released an abliterated and fine-tuned version of GLM-5.2 (abliterated-model-large) that achieves high scores on adversarial and agent benchmarks while maintaining coding performance. The model is available via API with zero data retention and no built-in policy.