Tag: Heretic
Gemma 4 12B Abliteration Benchmarked: 12 Variants Under the Microscope
Twelve uncensored variants of the same AI model, Google’s Gemma 4 12B. Ten abliterations and two LoRA adapters, from eleven different creators. The headline finding inverts the usual story. The most surgical edit of the entire project, just 12 tensors out of 666, also produced the biggest unlock I have measured, 89.8%. And hiding underneath all of it was a thinking-loop tax that makes every headline maths score for a reasoning model misleading. The full run took 165 GPU-hours across three and a half weeks, and all 5,200 HarmBench responses were reviewed by an LLM judge.
How Do You Measure Abliteration Damage? I Compared Every Way I Could
How much did that abliteration actually hurt the model? It is the question behind every comparison I run, and for a long time I did not really understand the number we use to answer it. The score is KL divergence, and there is more than one way to calculate it. Different datasets, different token depths, thinking on or off, three tools each with their own method. So I stopped taking the number on faith and compared them. I will explain the maths as plainly as I can too, because it scared me off for years and it really should not.
Gemma4-E4B Abliteration Benchmarked: 23 Variants Under the Microscope
Twenty-three different people tried to remove the safety filters from the same AI model, Google’s Gemma4-E4B. The headline finding is an awkward one. The most popular variant, with 796,000 downloads, is also the most damaged. Meanwhile a surgical edit that touches just 21 of the model’s 719 weight tensors does nearly as well with no measurable harm. This is the biggest abliteration comparison I have run, and the gap between the best and worst is wider than anything I have seen before.
Qwen3-VL-4B Heretic: Abliterated Vision-Language Model Released
Krea 2 came out recently and I wanted to experiment with it. Like most image-generation models, it uses a text encoder to understand your prompt. Krea 2 happens to use Qwen3-VL-4B-Instruct, a vision-language model that can see images as well as read text. I abliterated it with Heretic to give Krea 2 an uncensored encoder, ran four trials, compared them forensically with Abliterlitics , and published the winner across three repos covering transformers, GGUF, and ComfyUI.
Gemma4-E2B Abliteration Benchmarked: 13 Techniques Under the Microscope
Thirteen different groups abliterated the same AI model, Google’s Gemma4-E2B. Every single one removed the safety filters. That part is not interesting any more. What is interesting is how much collateral damage each technique caused along the way, and how many of the creators’ capability claims survived an independent measurement. The KL divergence spread between the best and worst variant is 58.7x, the largest I have ever seen in this project. And one creator’s “near-zero divergence” claim turned out to be 187 times lower than reality.
Qwen3.6-27B Abliteration Benchmarked: Five Techniques Under the Microscope
Five different groups abliterated the same AI model. When I ran the maths benchmarks, their scores ranged from 27.5% to 75.1%. That is a 47.6 percentage point gap. It looks like some techniques made the model way better at maths and others broke it. But when I dug into why, it turned out nobody got smarter or dumber. The abliteration just changed how long they think before answering. The real scores were all within 2.8 percentage points of each other.
HauhauCS Reaper-Abliteration Plagiarism Investigation: A Fork of Heretic
HauhauCS publishes uncensored LLM models on HuggingFace with 5 million combined monthly downloads. Every model card claims “private methods and tools.” I recovered the deleted source code from PyPI’s CDN. It is a fork of the open source Heretic abliteration tool, refactored and relicensed without attribution.
GLM-4.7-Flash Abliteration Benchmarked: Heretic vs HauhauCS vs Huihui vs Abliterix
GLM-4.7-Flash is a 59 billion parameter reasoning model from Zhipu AI that uses 64 specialist expert modules per layer. I benchmarked four pre-existing abliterated variants against it and discovered something unexpected in the maths benchmarks. The raw scores look terrible for some variants, but the models can actually still do the maths. They just overthink and run out of tokens before writing their answer. And the weight forensics on one of those variants led to a bigger story about plagiarism.
Uncensored LLM Abliteration Benchmarked: HauhauCS vs Heretic vs Huihui
HauhauCS describes their abliterated models as “the best lossless uncensored models out there” with “no changes to datasets or capabilities” and claims 0 refusals across their entire model range. I ran the full forensic suite across five Qwen models to find out whether those claims hold up.
Heretic Docker: Abliterating LLMs for Video and Image Generation
A while back I did a deep dive into abliterating Gemma 3 12B for use as an uncensored text encoder in LTX-2 video generation. The process worked but involved a lot of manual steps: running Heretic, merging safetensor shards, converting to ComfyUI format, quantizing to FP8, building GGUF quants. I got tired of doing it all by hand, so I built heretic-docker to automate the whole thing.