Edited Qwen Model Slashes Refusal Rates Without Training

Developer Jonathan Coletti has released Qwen3.8-27B-Uncensored, an "abliterated" variant of Alibaba's 27-billion-parameter multimodal model. By applying direct weight edits to the base checkpoint rather than traditional fine-tuning, Coletti successfully reduced safety refusals on a set of 100 held-out harmful prompts from 98 to just 12. This dramatic shift in behavior was achieved with a minimal performance penalty, as the model experienced a mean loss of only 0.5 points across four multiple-choice benchmarks.

The modification was performed using Heretic, an automated abliteration tool based on 2024 research by Arditi and colleagues. Heretic uses Optuna's tree-structured Parzen estimator to search for optimal parameter combinations. The process targets the residual stream, a shared activation pathway in transformer models that carries information between blocks. By identifying and removing the specific activation direction associated with refusal behavior, the tool directly edits the attn.o_proj and mlp.down_proj weights. This mathematical approach requires no training data, gradient updates, fine-tuning, or reinforcement learning from human feedback, leaving the vision tower completely untouched.

For practitioners, this development highlights a highly efficient alternative to traditional model alignment and steering. Instead of spending massive computational resources on retraining, developers can surgically alter model behavior. The resulting model is released under an Apache 2.0 license. Running the model in bf16 precision requires approximately 55GB of VRAM, and the release includes GGUF quantizations as well as support for vLLM and SGLang serving. Coletti also successfully grafted the model's MTP speculative-decoding head back onto the architecture after a standard transformers re-save had dropped it.