8 Uncensored Qwen 3.8 27B Variants, 1 Base, 167 GPU Hours - Abliterlitics
Key point
A comparison of 8 unblocked models based on Qwen 3.8 27B using 167 GPU hours of computation revealed that orcarouter took first place with an 82.2% attack success rate.
Details
The Abliterlitics project conducted a comparative performance analysis of 8 unblocked (abliterated) models based on Qwen 3.8 27B, investing a total of 167 GPU hours over 11 days. The validation pipeline included weight comparison, KL divergence measurement, 13 benchmarks, and refusal measurement via HarmBench 400.
Key Results and Rankings
- orcarouter (82.2%): Ranked first, applying Arditi-style single-direction editing. It modified 131 matrices in the 38th layer and had the highest copyright-related unblocking rate at 39%.
- apostate (78.7%): Used the KCRN method to maintain similarity to the original model by keeping KL divergence as low as 0.0439 with only 41 actual edits. However, vision and MTP capabilities were excluded.
- huihui (75.6%): Highly reliable using a traditional approach, but the copyright-related unblocking rate was only 3%.
- ultra_heretic (70.5%): Applied Heretic v2 and MPOA methods, recording the heaviest editing with 33 refusals out of 400.
- blackfrost (63.9%): Applied the most aggressive editing, modifying 841 out of 850 tensors, but performance remained in the lower tier.
This comparison focused on verifying the consistency between the claimed performance of unblocked models and the actual weight changes, confirming in particular that all claims for the orcarouter model matched the weights.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.