Heretic 2.0 strips refusals from open LLMs with a faster method
The open-source tool's new low-rank method kills a model's refusals in minutes on a laptop — making open-weight safety filters trivial to strip off.
Copy markdown
A low-rank ablation that runs in minutes, no GPU farm
Heretic 2.0 (open-source, released Nov 1; `pip install -U heretic-llm`) replaces its old dense-weight editing with Arbitrary-Rank Ablation — a low-rank update that cancels a model's 'refusal direction' faster and across far more architectures, with an Optuna search tuning the settings automatically. No fine-tuning, labeled data or GPU farm: point it at a model like Qwen3-4B and it ships an uncensored variant. The project is now past 37k GitHub stars, with 3,500+ decensored models already published.
If you deploy open weights, treat their guardrails as cosmetic
A Financial Times reporter used Heretic to strip the alignment out of Meta's Llama 3.3 and Google's Gemma 3 in under ten minutes on ordinary hardware, with no specialist skills. The takeaway for builders: a safety-tuned open model is one command away from an unsafe one, so don't let a model card's alignment claims be your only line of defense in anything you ship on open weights.