Qwen3.5-4B Heretic Slashes AI Refusals by 87% Without Breaking Benchmarks

A community-modified build of Qwen3.5-4B has been published on Hugging Face with sharply reduced refusal behavior and little reported benchmark drift. According to the model card , Qwen3.5-4B-heretic lowered refusals from 99 to 13 on a 100-prompt test. Its smallest listed GGUF quantization is a 2.71 GB file suitable for local inference on a laptop.

The release also demonstrates how Heretic 2.0 has automated a model-editing technique commonly called abliteration. The pip-installable tool analyzes supported transformer models, identifies activation patterns associated with refusals, and modifies the weights without conventional fine-tuning. The release reports a runtime of about 45 minutes, though actual time depends on the model, GPU, configuration, and evaluation workload.

Refusal behavior, mapped into the weights

Abliteration edits existing weights by estimating directions in activation space that correlate with refusal responses. The process uses contrasting prompt sets to locate those patterns, then suppresses their influence while attempting to preserve responses on ordinary tasks. This approach targets a measurable component of safety tuning; it does not establish that every refusal mechanism or safety behavior resides in one direction.

This build uses Arbitrary-Rank Ablation, or ARA, which learns a low-rank subspace across multiple layers. Its optimization combines behavior preservation with stronger movement away from refusal patterns. Heretic modified layers 15 through 27 of the language tower, using a preserve-good-behavior weight of 0.99 and an overcorrection weight of 0.82. The vision encoder remained unchanged during the edit.