Price terms as published
Each line is the vendor’s own wording, split from the price capture checked on 2026-09-21. Nothing here is rounded, converted or re-worded — the full sentences stay in the review below. Lines that carry only a price continue the plan named on the line above, in the vendor’s own order.
| # | Plan / term | Checked |
|---|---|---|
| 1 | Free (open source, AGPL-3.0) | 2026-09-21 |
One-liner: Heretic is a command-line tool that removes “safety alignment” (censorship) from transformer language models automatically, without post-training, by ablating a small number of directions in the model.
Verified facts (as of 2026-09-21)
| Attribute | Value | Source |
|---|---|---|
| GitHub stars | 31,997 | repo page snapshot, 2026-09-21 |
| Forks | 3,596 | repo page snapshot |
| License | AGPL-3.0 | repo license field (spdxId verified on page) |
| Last commit | 2026-09-05 | commits feed |
| Method | Directional ablation (“abliteration”, Arditi et al. 2024) + TPE parameter optimizer built on Optuna | repo README |
| Objective | Co-minimize refusals and KL divergence from the original model | repo README |
| Model support | Most dense models, many multimodal, several MoE architectures, hybrid models like Qwen3.5; pure state-space models not supported out of the box | repo README |
| Mirrors | Codeberg mirror + Hugging Face org (“heretic-org”) | repo README badges |
How the automation works
Earlier abliteration recipes required hand-tuned choices about which residual directions to ablate. Heretic’s contribution is search: an Optuna-driven TPE optimizer explores parameter combinations, scoring candidates by how fewer refusals they produce while staying close to the original model’s distributions (KL divergence). The README’s positioning is that anyone who can run a CLI program can decensor a model with it — no transformer internals knowledge needed.
Two red lines to state plainly
- License: AGPL-3.0, copyleft. The repository’s own license field is AGPL-3.0. Linking or embedding Heretic into a network service triggers source-disclosure obligations for that service. Anyone writing this up as “MIT-style, safe for commercial embedding” is wrong. Also note that whatever base model and weights you feed it carry their own licenses — ablating a model does not launder its terms.
- Use-case risk is yours. The tool’s stated purpose is removing censorship from language models. Distributing or deploying decensored models can violate upstream model licenses, hosting platforms’ policies, and in some jurisdictions legal requirements. This page documents what the software does; it is not an endorsement of any particular deployment.
Who should use it
Researchers and hobbyists evaluating refusal behavior on open-weight models they have the license to modify, on infrastructure where AGPL compliance and distribution policy are already solved. Teams looking for a compliant production pipeline should treat both red lines as blocking until legal has reviewed.
Stats, license field, method description, and support matrix verified against primary sources (GitHub repo page, raw README, commits feed) on 2026-09-21. Official-sources-only entry: no decensoring run was performed here, and no quality comparison against manual methods is claimed.