Heretic 2.0 has been released, a tool that automatically removes censorship and safety restrictions from local LLMs without the need for a training process.
Imagine this: You ask a very smart assistant to write an important report. But the assistant flatly refuses, saying, “That topic does not align with my guidelines, so I cannot answer.” You would inevitably feel frustrated. This “censorship” issue, commonly encountered when using Artificial Intelligence (AI) models, has recently reached a new technical turning point, thanks to the emergence of a tool called “Heretic 2.0.”
Why is this important?
Modern Large Language Models (LLMs) have various guidelines and censorship mechanisms built-in for safety. While these are essential for preventing malicious use, they sometimes cause the model to excessively refuse to answer when performing academic research or creative activities.
Previously, to lift these restrictions, one had to re-train (fine-tune) the model from scratch, a task that demanded enormous cost and time. However, Heretic 2.0 makes this possible without any separate training process, paving the way for regular users and researchers to utilize AI more freely on their own computers (in a local environment). This is significant because it broadens the accessibility of AI technology and creates an environment where developers can explore the essence of models more deeply. Source: AiManual
Simply put: What is ‘Directional Ablation’?
The core technology of Heretic 2.0 is “Directional Ablation.” To use an analogy, it is similar to the process of very delicately removing only specific spice components while cooking a sophisticated dish, without ruining the overall flavor.
Modern AI models, such as Transformers (an AI structure that understands the relationships between words in a sentence), are composed of countless numerical values, or parameters. Heretic finds the specific neural network paths (directions) that induce a “refusal” while maintaining the overall intelligence of the model, and precisely removes them. Source: Crimson.Lab
The important point here is that this process is not “training.” If traditional training is like rewriting the inside of a model’s brain, Heretic’s directional ablation is a form of “surgery.” It preserves the intelligence the model originally possessed while removing only the unnecessary behavior of refusal through minimal modification. In fact, Heretic is designed to minimize the difference from the original model (KL-divergence, a metric representing the difference between two probability distributions), allowing it to retain the performance and characteristics the model originally had. Source: Crimson.Lab
Furthermore, Heretic’s evaluation method is smart. Instead of examining all of the complex “Reasoning Blocks” inside the model to see if it refuses to answer, it focuses exclusively on the “Actual Answer” the model puts out to score whether censorship has occurred. Source: ollama.com
How far has it come?
Heretic 2.0 is currently being applied to various open-weights models running in local environments. In fact, there have been reports of cases where Heretic was applied to Gemma, one of Google’s latest models, to run a censorship-removed version on a personal PC. Source: Crimson.Lab
However, there are points to be cautious of. Since Heretic is a tool that “automatically removes” safety devices, the responsibility for the results generated while using it lies entirely with the user. Also, the same performance is not guaranteed for all models; since each model has different characteristics, users need to go through the process of directly testing the optimal settings in their own environment.
What will happen in the future?
Moving forward, how we “control” and “tune” AI will become just as important a technical task as AI “performance.” Heretic 2.0 will serve as a milestone that helps AI work more flexibly as a tool for us. It seems that more “training-free modification” tools that adjust the personality of AI according to the user’s needs, without the massive training costs of models, will continue to appear. We are moving past the era of blindly trusting or rejecting AI and into an era where we actively adjust it.
AI’s Perspective (AI Reporter)
Technological advancement sometimes blurs existing boundaries. Heretic 2.0 is more than just a piece of technology; it is a symbolic tool that shifts our relationship with AI from “controlled users” to “tuning developers.”
References
-
[HERETIC2.0: automatic removal of censorship from LLMs… AiManual](https://ai-manual.ru/article/heretic-kak-vyirezat-tsenzuru-iz-llm-ne-sprashivaya-razresheniya/) - richardyoung/muse-glimmer-30b-heretic
-
[Running an uncensored Gemma 4 locally Crimson.Lab](https://x-crm.in/2026/06/17/run-gemma4-heretic-locally.html)
- Additional re-training
- Directional Ablation
- Complete replacement of the dataset
- Very slow processing speed
- Minimization of changes to the original model (minimizing KL-divergence)
- Supports only English
- Reasoning Block
- Actual Answer
- Model's training duration