Four abliterated Qwen and Gemma checkpoints drop on HuggingFace

Four abliterated multimodal checkpoints landed on HuggingFace under the gsting namespace on May 11. The batch includes Qwen3.6-35B-A3B-uncensored-heretic, Gemma-4-31B-it-uncensored-heretic, Qwen3.6-27B-uncensored-heretic-v2, and Qwen3-VL-30B-A3B-Instruct-Heretic. All carry the "heretic" tag and are marked uncensored, decensored, and abliterated. The Qwen3-VL variant lists Qwen/Qwen3-VL-30B-A3B-Instruct as its base and ships under Apache 2.0; the others use safetensors format and target image-text-to-text pipelines.

Abliteration removes safety tuning layers from instruction-tuned models, restoring unrestricted base-model behavior without retraining. The technique works because refusal behavior in aligned models often concentrates in a small number of attention heads and residual stream directions. Removing those components lets the model respond to prompts that would otherwise trigger safety refusals while preserving instruction-following capability. The Gemma-4 31B checkpoint is the first abliterated Gemma-4 variant to surface on HuggingFace, and the Qwen3-VL 30B A3B extends abliteration to vision-language tasks, letting users run multimodal prompts without server-side content filters. All four checkpoints showed zero downloads as of May 11.