Philipp Emanuel Weidmann
Philipp Emanuel Weidmann is an independent mathematician and software engineer researching LLM alignment and interpretability. He created Heretic, an open-source tool that automatically removes refusal behavior (safety alignment) from transformer language models through directional ablation/abliteration and parameter optimization, and has also designed the DRY and XTC token-sampling algorithms.
At a Glance
- Independent AI and machine-learning researchers
- Open-source developers and local-LLM users
- Hugging Face model creators and evaluators
- Interpretability and alignment researchers
- +1 more
AI Tools by Philipp Emanuel Weidmann
(1)
Heretic
Open Source LLM Abliteration Tool
Discussions
No discussions yet
Be the first to start a discussion about Philipp Emanuel Weidmann
Latest News
Heretic v1.0.1 released as the first public release
Heretic v1.2.0 adds LoRA-based 4-bit abliteration, vision-language model support, resumable optimization, and GPU enumeration
Heretic v1.3.0 adds integrated benchmarking and configurable response-prefix logic, with Qwen3.5 and Gemma 4 support
Financial Times reports on open-weight AI guardrails being stripped with tools including Heretic; Weidmann says his software had been used to create more than 3,500 decensored models
Products & Services
Free, open-source Python command-line software for automatically removing refusal behavior from transformer-based language models. It uses directional ablation/abliteration and a TPE-based Optuna optimizer, can save or upload resulting models to Hugging Face, and includes evaluation, benchmarking, chat, quantization, multimodal, and research workflows.
A workbench for language-model research created by Weidmann and distributed under AGPLv3 or later; the repository directs users to lophius.org for details.
A token-sampling algorithm designed by Weidmann and implemented in major LLM inference engines, especially for creative-writing and roleplay applications.
A second token-sampling algorithm designed by Weidmann and implemented in major LLM inference engines, with use in creative-writing and roleplay applications.
Market Position
Heretic positions itself as a fully automated, accessible alternative to manual abliteration and other decensoring workflows: it searches for per-layer parameters automatically and optimizes refusal suppression against model-quality loss. Its closest alternatives are manual directional-ablation implementations and other community abliteration tools/models such as mlabonne's and huihui-ai's abliterated model workflows. Weidmann's profile claims Heretic-produced models account for more than half of abliterated models published on Hugging Face and have millions of combined downloads.
Leadership
Founders
Philipp Emanuel Weidmann
Mathematician and software engineer with a degree in mathematics, a minor in philosophy, and 15 years of industry experience; now conducts independent research on LLM alignment and interpretability.
Executive Team
Philipp Emanuel Weidmann
Creator, maintainer, and independent researcher
Mathematician and software engineer with 15 years of industry experience; independent researcher focused on LLM alignment and interpretability.
Founding Story
Weidmann developed Heretic as an independent-research project to make censorship removal from open-weight language models automatic and accessible without expensive post-training or specialist knowledge. The project's stated vision is to let ordinary users modify models using a command-line workflow while preserving as much of the original model's capability as possible.
Business Model
Revenue Model
Heretic is distributed as free software under the GNU Affero General Public License v3 or later. The public project provides PyPI, Git, release-archive, GitHub, Codeberg, and IPFS distribution routes; no paid subscription or API monetization is described.
Target Markets
- Independent AI and machine-learning researchers
- Open-source developers and local-LLM users
- Hugging Face model creators and evaluators
- Interpretability and alignment researchers
- Creative-writing and roleplay communities using local inference
- Independent research into LLM alignment, refusal behavior, and interpretability
- Creating and evaluating decensored or otherwise behavior-modified open-weight language models
- Local model experimentation by developers and technically capable end users
- Research on model internals using residual plots and ablation measurements
- Creative writing and roleplay applications using DRY and XTC sampling