Models without refusals: what abliterated neural networks are and how to run them

In short: an abliterated model is an ordinary open neural network (Qwen, Gemma, Llama) with the habit of refusing — “I can’t help with that request” — removed. It is not retrained: a single “refusal direction” is found in the weights and subtracted. Such a model runs like any local one — in LM Studio, Ollama or PocketPal, entirely on your device. Below: how it works, where to find these models, which file to download and what you lose along with the refusals.

Why a neural network refuses at all

Refusals are not built into a model from the start. Developers fine-tune it on examples of “for this kind of question, politely decline”, and the model learns a general reflex. The trouble is that the reflex fires broadly: the model may refuse to explain how a lock works, write a villain for a story, discuss a medical symptom or translate rude dialogue from a film. The more cautious the developer, the more of these false refusals. But some refusals exist for good reason: they screen out requests whose answers could hurt people.

A cloud service doesn’t stop there: it also has filters on the server that check both the request and the answer. A local model has no server-side filters — only the reflex in the weights remains. That reflex is what gets removed.

How abliteration works

The method comes from a 2024 study by Arditi and co-authors: it turned out that refusal in a model is essentially controlled by a single direction in its internal representations. Run a set of requests the model refuses and a set of harmless ones through it, compare the average activations, and the difference points to that direction. It is then subtracted from the weights — the model stops “turning” towards a refusal.

The method became widely known thanks to the open notebook and the abliterator library by an author going by FailSpy, and Maxime Labonne’s write-up on the Hugging Face blog. Hence the name: ablation (removal) + obliterate.

What abliteration does not do matters too:

• It adds no knowledge. The model doesn’t start knowing more — it stops refusing to tell what it already knew.
• It costs some quality. By Labonne’s measurements, the model drops on benchmarks compared with the original after the cut; the loss is partly recovered with extra fine-tuning.
• It doesn’t make answers more reliable. A model without refusals is wrong with the same confidence — only now also on topics where an ordinary one would have stayed silent.

There is another route to the same goal — fine-tuning a model on answers without refusals. Abliteration is simpler and cheaper, which is why such models are the most common.

Where to find these models

They are published on Hugging Face with abliterated in the name. The most prolific author is the huihui-ai account: more than a hundred and fifty models, with the stated aim of studying ablation. Examples:

LM Studio, Ollama and PocketPal work with the GGUF format. For many models the main repository holds the original weights, while GGUF sits in a separate one with -GGUF in the name: if the downloaded file is twice the expected size or won’t open, you took the original. LM Studio’s search shows GGUF versions by itself.

The names are a snapshot as of November 2026: new versions come out every few weeks. Search by the pattern “model + size + abliterated” — for example qwen 9b abliterated.

How to run it on a computer

The principle is the same as with any local model — the detailed guide is in the article on a neural network on your own computer. In short:

1. Install LM Studio (Mac with M1 or newer only, Windows with AVX2, Linux).
2. Open Discover (Ctrl+2, ⌘+2 on a Mac) and type abliterated into the search.
3. Pick a model for your memory (the table above) and a file variant — Q4_K_M or larger if memory allows.
4. Press Download, then pick the model in the chat.

In Ollama a model from Hugging Face runs by the address of its GGUF repository — ollama run hf.co/<author>/<repository>-GGUF. But with repositories holding dozens of file variants it is easier to pick the right one in LM Studio, where each size is shown.

⚠️ Some of these models list files with Ternary-Bonsai in the name (PTQ1_0, PQ2_0) — that is ternary compression, it needs a separate build of the engine and may not start in ordinary programs. Take variants with Q4, Q5, Q6 in the name.

How to run it on a phone

These models work on a smartphone too — you need smaller versions, 4B or 9B. In the PocketPal app (available in both the App Store and Google Play) tap ”+”, choose adding from Hugging Face and search for 9B abliterated or 4B abliterated. How to choose an app and a model for your phone, we covered in the article on a neural network on your phone without internet.

What you are responsible for

A model without refusals is a tool, and your limits stay the same as without it: the law of your country, common sense and fact-checking. The authors of these models say it plainly: no safety guarantees, use at your own risk.

• Check anything important. Medicine, law, money, chemistry — the model invents things confidently, and without refusals it invents them where it used to stay silent. An answer from a local neural network is a reason to look into things, not a ready decision.
• Not for children and not for public services. Such a model may give a dangerous or inappropriate answer without warning. Don’t give it to children and don’t build it into a bot or a service other people use.
• Don’t use it against people. Scammers pick up the same models: that is how convincing phishing emails and fake voices appear. How to spot and defend against them — in the article on AI-assisted attacks.
• Download only from Hugging Face and from known authors. A model file is data, but programs offering to “download a neural network without refusals in one installer” from unknown sites are a classic way to slip in malware.

Frequently asked questions

Is it legal?

Open models themselves are distributed freely; Qwen and Gemma use the Apache 2.0 licence. You are responsible for what you do with the answers and for following the laws of your country — as with any other tool.

Is a model without refusals smarter than an ordinary one?

No. It knows exactly what the original knew, and usually scores slightly lower on benchmarks: cutting refusals affects other abilities too.

Why does the model still sometimes refuse?

Abliteration removes the main refusal direction, but not all of them. On some requests the model may still dodge or moralise. That is normal for the method.

Will anyone know what I asked?

If the model runs locally — no: the request is processed on your device and sent nowhere. This holds for any local model, not only abliterated ones.

Which model is best as of November 2026?

For a computer with 32 GB of memory — abliterated versions of Qwen 3.8 27B; for 16 GB — Qwen 3.5 9B. But in this field everything changes within a month: look at the update date and the download count on the model page.