Heretic – v1.2.0
🚨 Heretic v1.2.0 is live — and it’s packing serious upgrades! 🚨
The team just dropped a massive update to Heretic, the fully automatic censorship-removal tool for LLMs — and this one’s packed with performance boosts, new features, and rock-solid stability fixes. Here’s what’s new:
🔹 Memory & Stability Wins
- `amax_memory` setting to cap RAM usage 🧠
- Smarter iteration logic to avoid getting stuck in low-divergence traps ⚙️
- Magnitude-Preserving Orthogonal Ablation — finer control over refusal suppression 🎯
🔹 LoRA + Quantization Power-Ups
- Brand-new LoRA-based abliteration engine 🪄
- Full 4-bit quantization support — run decensoring on consumer GPUs 🖥️
- Auto-GPU detection at startup (no more manual config headaches!) 🤖
🔹 Resilience & Flexibility
- Run extra optimization trials after the main run finishes 🔄
- Save/resume support for long jobs — no more lost progress! 💾
- Fixed MXFP4 loading issues 🛠️
- Refactored save/load machinery = way more reliable 📦
🔹 Vision-Language Models? Now Supported!
- Broad support for VL models (e.g., LLaVA-style) 📸➡️🧠
- Plus: full type checking, debug tools, slop-reduction configs, and bug fixes 🐞
🌟 Shoutout to @noctrex, @accemlcc, @anrp, and @salmanmkc — welcome to the crew!
🔗 Full changelog: v1.1.0 → v1.2.0
🚀 Grab it now and keep decensoring smarter, faster, and safer!