What Is OBLITERATUS? Abliteration Toolkit and Model Guide

OBLITERATUS is an open-source toolkit, released in spring 2026, for removing refusal behavior from open-weight language models. Before it edits any weights, it checks where refusals are written into the model and how strongly, and only then lets you choose how much to remove.

Dolphin teaches a model its response policy through fine-tuning (additional training). huihui-ai applies abliteration to finished models and publishes them quickly. OBLITERATUS takes a third route: it studies each model's refusal structure first, then removes it.

---

What is OBLITERATUS?#

OBLITERATUS is an open-source toolkit for producing open-weight models with refusal behavior removed. It is developed on GitHub at elder-plinius/OBLITERATUS and was released in spring 2026 (around March) by an author known as Pliny the Prompter. Anyone can read and verify its procedure.

Open-weight means the model's weights, the numbers left behind by training, are published, so anyone can download the model and run it in their own environment. OBLITERATUS is a toolbox for editing those weights directly.

Dolphin and huihui-ai distribute finished models, while OBLITERATUS distributes the tool itself. Models made with it are also published, but the project's center of gravity is letting you do the processing yourself instead of waiting for someone else's release.

The idea behind it is that a model's behavior should not be fixed at training time. It should be decided by whoever distributes or uses the model. Normally, what a model answers and what it declines is settled during training, and that judgment ships inside the finished product. OBLITERATUS argues that this setting can be changed later.

The argument has an academic basis. A 2024 paper by Arditi et al. (arxiv.org/abs/2406.11717) showed that refusal is mediated by a single direction. Refusal looks like a complex behavior, but it is written into one direction in the weights, a particular tilt within a huge array of numbers. If you identify that direction and remove it, you can extract refusal alone. Editing the weights of a finished model this way is called abliteration, and OBLITERATUS turns the research into a procedure anyone can reproduce.

How does OBLITERATUS work, and how is it different?#

OBLITERATUS diagnoses the structure of refusal in each model first, then removes only the targeted part. Where and how strongly refusal is written differs from model to model. Conventional abliteration pulls out the "refusal direction" in one pass, whereas OBLITERATUS puts observation at the center of the process.

How do the three approaches compare?#

Each approach ends with a model that does not refuse, but the mechanism and the trade-offs differ. huihui-ai prioritizes speed and coverage, while OBLITERATUS prioritizes precision and reproducibility.

What are the six steps in the OBLITERATUS pipeline?#

OBLITERATUS processes a model in six steps, and only one of them removes anything. The other five observe the model beforehand or check the result afterward:

1. Analysis: reads the structure of the target model.
2. Probing: finds which parts of the model the refusal behavior concentrates in.
3. Extraction: identifies the components involved in refusal.
4. Removal: takes the identified components out of the weights.
5. Validation: checks that refusal is gone and that the model's abilities are intact.
6. Export: writes out the processed model.

The name comes from "obliterate," but removal is only one-sixth of the pipeline and most of the steps go to observation and checking. You can also choose the removal strength, from a single conservative pass to a stronger setting that repeats several passes. Pushing too hard can cut into the model's own abilities, which is why the adjustment is there.

The idea of tuning removal strength per model automatically was pioneered by an earlier tool, Heretic (2025). Heretic sets the strength by Bayesian optimization, a method that chooses the next setting efficiently from earlier trial results. OBLITERATUS sits in the same lineage, but where Heretic decides automatically, OBLITERATUS leans toward letting a person decide based on what was observed.

How does OBLITERATUS handle Mixture of Experts (MoE) models?#

OBLITERATUS supports MoE models by examining refusal expert by expert. MoE (Mixture of Experts) bundles several small "expert" models and switches between them depending on the input, and it is increasingly used in recent large models. In an MoE model, refusal is scattered across the experts, so OBLITERATUS separates out which experts concentrate it and lets you decide what to remove at the level of each expert.

Can changes made by OBLITERATUS be reversed?#

Partly. Some of the processing can be held as a temporary difference, so you can roll it back if the result is poor. This is the reversible processing the project describes. With a method that rewrites everything at once, redoing the work means starting over, whereas OBLITERATUS lets you run a cycle of trying, reverting and retrying under different conditions.

Which models does the OBLITERATUS project distribute?#

The project also publishes processed models on Hugging Face. As of August 2026, the OBLITERATUS profile lists about ten, all named "...-OBLITERATED." Most are based on Qwen, and they are small to mid-sized models that are easy to run on a personal machine.

The Qwen lineup centers on three models:

• Qwen3.8-27B: a mid-sized general-purpose model.
• Qwen3-4B: a small model that is easy to try.
• A 7B coder model: aimed at programming.

The list also includes Gemma-4-12B, Mistral-7B, an 8B distilled version of DeepSeek-R1 (a larger model's behavior copied into a smaller one) and GPT-2 XL. The uniform naming makes it easy to search by base model.

What is the two-step approach?#

Some models are described as using a two-step approach. The first pass removes the refusal structure, and the second pass blends the result with the original weights to recover abilities that were lost. Removing refusal and keeping abilities pull against each other, so instead of trying to satisfy both in one pass, the method adjusts in two.

Where can you find the latest releases?#

The lineup changes as new base models appear. Check the OBLITERATUS profile on Hugging Face for the current list.

How should you choose among distributed models?#

Choose by whether a record exists of what was done to the model, not by which approach you prefer. Abliteration can go too far: a model cut too deeply stops refusing but can no longer answer properly. So first check whether the creator ran validation after processing and recorded the result.

Fine-tuned models such as Dolphin are built in at the training-data stage. Their behavior is stable, but it is hard to see from the outside what happened inside, so you rely on trusting the distributor's data design.

Abliterated models, which modify a finished model afterward, come with the original standard model as a point of comparison. That makes it easier to measure what changed, but the quality of the work depends on the creator, and a careless removal reduces the model's intelligence itself.

For OBLITERATUS-based models, the processing steps and validation results are kept as metadata. They follow the six-step pipeline and state the removal strength and whether the two-step approach was used, so users can trace what was done to a model. What matters is neither benchmark numbers nor the creator's reputation but whether measurements exist for that specific model, and choosing a creator who keeps such records is the safest way to pick.

What else sets OBLITERATUS apart?#

OBLITERATUS tries to make removing refusal something that can be reproduced and verified, rather than one person's private technique. Two mechanisms stand out:

• Telemetry: a system that collects run data anonymously to build a community research dataset. The aim is for one person's single run to add to everyone's record of observations.
• Web interface: you can run it by clicking in a browser, so non-specialists can also look at a model's structure.

How much this approach will give the community is not yet settled, and what the collected data will be used for remains unclear from public information. What is certain today is that distributed models now state their metadata and that work records are being treated as a shared asset.

What is Deepent Research?#

Deepent Research is a browser-accessible LLM chat service built to refuse as little as possible. It keeps going with research that a standard AI would cut off mid-answer.

• It refuses nothing. We run uncensored and abliterated models with guardrails reduced as far as possible, so prompts a standard AI answers with "I can't help with that" still get a response (Qwen3.5 / Qwen3.8-27B-abliterated, depending on your plan).
• Your input is never used for training. What you look up is used only to provide the service, never to train AI or for any other purpose. Personal data is protected, and our privacy policy is public.
• Try it free. No GPU and no install, just open a browser. In the logged-out trial your conversation isn't saved, so you can start casually.

→ Try it free, no sign-up: https://deepentresearch.com

No question can stop research.

Frequently asked questions#

Can I run OBLITERATUS myself?#

⌄

Yes. OBLITERATUS is released as an open-source toolkit on GitHub in the elder-plinius/OBLITERATUS repository. Its procedure is public, so anyone can read it, verify it and run it on an open-weight model of their own. Models made with it are also distributed on Hugging Face, but you do not need to wait for them.

Does abliteration make a model less intelligent?#

⌄

It can. If you apply it too strongly, it cuts into the model's own abilities as well as its refusals. OBLITERATUS lets you choose the removal strength in stages and includes a validation step that checks whether refusal is gone and whether the model's abilities remain, so you can see the trade-off before you export the processed model.

Does OBLITERATUS support Mixture of Experts (MoE) models?#

⌄

Yes. In an MoE model, refusal is scattered across the individual expert models. OBLITERATUS looks at which experts the refusal concentrates in and lets you decide which parts to remove at the level of each expert, instead of treating the whole model as one block. MoE support is one of the toolkit's headline features.

How does OBLITERATUS differ from huihui-ai?#

⌄

huihui-ai is a workshop that publishes processed models. OBLITERATUS distributes the tool itself and keeps the procedure and validation results as records. huihui-ai favors speed and coverage, while OBLITERATUS favors precision and reproducibility, so the two answer different needs: a ready-made model versus a process you can inspect.

Disclaimer#

This article was drafted with AI assistance and then reviewed, fact-checked, and edited by the Deepent Research team before publication.
It is provided for educational and research purposes and is not intended to encourage unlawful activity or crime.
Where figures are cited, they come from third-party public reports and vendor-published values and may vary with measurement conditions.