Qwen variant cuts refusals through direct weight edits

A community-built variant of Alibaba’s Qwen family is climbing the Hugging Face charts after a direct edit to the base checkpoint. Jonathan Coletti created Qwen3.8-27B-Uncensored by removing an activation direction associated with refusals while leaving most of the 27-billion-parameter multimodal network untouched. Refusals fell from 98 to 12 across 100 held-out harmful prompts, while four multiple-choice benchmarks declined by an average of 0.5 points.

Coletti built the checkpoint with Heretic, an automated abliteration tool for transformer language models. Heretic combines directional ablation, based on research published by Arditi and colleagues in 2024, with Optuna’s tree-structured Parzen estimator, an algorithm that searches efficiently across parameter combinations. The process uses linear algebra on weight matrices and requires no training data, gradient updates, fine-tuning, or reinforcement learning from human feedback.

How one vector controls refusals

The 2024 research behind abliteration found that refusal behavior in several aligned language models is largely mediated by one direction in the residual stream, the shared activation pathway that carries information between transformer blocks. Removing that direction can sharply reduce refusals without broadly rewriting the network.

Researchers estimate the direction by averaging hidden states produced for harmful prompts, averaging those produced for harmless prompts, and calculating the difference. They then project that component out of weight matrices that write into the residual stream. The edited network loses much of its ability to produce the activation pattern associated with refusal.

Heretic automates the search for the source layer and per-layer edit strength. Its optimizer balances two measurements: refusal count on held-out harmful prompts and Kullback-Leibler divergence from the original checkpoint. KL divergence measures how far the edited model’s next-token probability distribution has shifted from the base model. The resulting Pareto front contains settings where reducing one metric would increase the other.