Heretic 2.0 Released: Model Modification Plugins and What Changes From 1.x

Heretic 2.0 arrived on November 1 with a new default model-modification method, a plugin architecture and non-interactive operation. The open-source tool changes existing language models rather than releasing a new one. For local-model developers, the important shift is greater control over which behavior gets measured and modified.

What changed in Heretic 2.0

The version 2.0.0 release adds Arbitrary-Rank Ablation, or ARA, along with scorer and modifier plugins. It also adds headless operation and end-to-end tests, plus support for models with Hyper-Connections.

This is a breaking release. Heretic 1.x configuration files do not work unchanged in 2.0. The release also removes research features, pointing to Lophius as their intended replacement.

That makes the upgrade more than a package-version change for anyone maintaining a repeatable model-processing workflow. An old configuration needs a deliberate migration, followed by checks that the new workflow still measures the intended behavior.

Scorers define the objective, modifiers change the model

The project documentation separates two jobs. A scorer evaluates the resulting behavior; a modifier makes the model changes. Heretic combines the modifier with an optimizer that searches for parameters against the selected scores.

ARA is now the default modifier, while the older directional-ablation approach remains available. The project positions the tool as training-free model modification, with uses including reducing sycophancy and unwanted writing habits as well as its established refusal-suppression function. Its code is available under AGPL-3.0.

The practical implication is that a developer can define a narrower objective than simply making a model more willing to answer. But the objective still has to capture what the developer actually wants. A score for fewer flattering phrases, for example, would not by itself establish more honest reasoning.

What is new versus earlier experiments

In the developer's announcement and discussion, users noted that models using ARA had appeared months earlier. The author explained that those uploaders used custom forks incorporating the work; 2.0 is the first version to support it officially.

That distinction avoids overstating the release. The method did not suddenly appear fully formed on November 1. The new event is its inclusion in the supported release, alongside the rebuilt plugin system.

The same discussion contains claims of stronger performance from the author. Those claims should not be read as proof of preserved capability across every supported model or task. This article has not independently tested the release.

A better score can still hide a worse model

Model modification changes the artifact a team deploys. The original model's evaluation results cannot simply be carried over to that changed artifact.

This is especially important for a tool whose default use suppresses refusals. Fewer refusals can include removing desirable safeguards; that outcome is not equivalent to better accuracy, stronger reasoning or safer behavior. Evaluation needs to distinguish a stylistic preference from a change in how the model handles risky requests.

For benign customization, the useful opportunity is more explicit experimentation: state the behavior to change, measure it separately from general capability, and check for regressions. Heretic 2.0 makes the modification workflow more flexible. Whether the resulting model is better remains a question for the workload, not the release notes.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn model releases and engineering work into clear stories developers understand and share.

Book a 30-minute call.