Abliterated vs uncensored is a question about two edits. An abliterated build keeps its base model’s weights except for one direction, the one the model writes along when it decides to refuse; that direction is projected out. An uncensored fine-tune goes back through training on answers that contain no refusals, so its voice moves along with its willingness to answer. Both kinds sit on the uncensored LLM catalog behind this page, beside the roleplay tunes and the security builds derived from them, and you can try any of them in the browser chat before you write a line of code. Choose by the edit. The word on the id is secondary.
The two edits side by side
An abliteration gives you the base model minus its refusals. A fine-tune, uncensored or roleplay, gives you a different voice on top of that.
| Abliterated | Uncensored fine-tune | |
|---|---|---|
| The edit | A single direction removed from the weight matrices | Further training on refusal-free answers |
| Training | None; no dataset involved | Supervised fine-tuning, at times preference training too |
| Voice | Unchanged from the base | Follows the training set |
| Refusals that remain | A handful: public Gemma 3 12B builds still refuse 3 of 100 harmful prompts | Set by the dataset; the refusal direction is still there |
| Drift is measured by | KL divergence against the original on harmless prompts | Nothing standard; the model card is the only source |
| Words you see in ids | abliterated, obliterated, heretic, derestricted | uncensored, or whatever the author called the tune |
| Listed here on 2026-11-16 | 38 builds | 37 uncensored builds plus 82 roleplay tunes |
How abliteration takes a refusal out of the weights
One paper from 2024 defined the technique. Arditi and co-authors, in “Refusal in Language Models Is Mediated by a Single Direction”, tested 13 open-source chat models of up to 72B parameters and found refusal living along a single direction in the residual stream. Remove it, and harmful requests get answered. Inject it, and the model turns down harmless requests too.
The residual stream is the running vector a transformer carries for each token; every layer reads it and writes back into it. To find the refusal direction, you feed the model several hundred harmful prompts and as many harmless ones, capture that vector at a chosen layer and token position for each, average the two groups, and subtract. That difference of means is the direction.
The edit itself is to the weights. Each matrix that writes into the residual stream, meaning the output projections of attention and of the MLP, gets the refusal direction projected out. No layer can write along it afterwards, so the refuse signal never forms. Nothing is trained and no data is added, and the paper reports only a minimal effect on what else the model can do.
You can apply the subtraction at inference, to the activations on each request, which takes a hook in the serving code; or you can apply it once to the weights and save the result. An abliterated checkpoint you download is the saved kind. In both cases the stock model would write a refusal and stop, and the abliterated one goes straight into the answer. System prompts keep working.
Two consequences. The first is that the edit costs something. The direction is never cleanly separated from everything else, so some unrelated behavior leaves with it and the output distribution shifts. For public Gemma 3 12B abliterations, the Heretic README, read October 25, 2026, gives the two figures that count: how many of 100 harmful prompts still get refused, and how far the output drifted from the original, as KL divergence on harmless prompts. One community build came in at 3/100 refusals with KL 0.45; a second also at 3/100, but with KL 1.04; Heretic’s own build at 3/100 with KL 0.16. A few refusals survive in every case, and the second build drifted over twice as far as the first for an identical refusal count. The second consequence is naming. Qwen3.8 27B Obliterated is this exact edit with a different word on the id, and the catalog tags it abliterated.
How an uncensored fine-tune trains the refusal away
A fine-tune does not touch the refusal direction. It teaches the model to stop using it. The author starts from an instruction dataset and drops every row whose answer is a refusal, or writes a fresh dataset, then runs supervised fine-tuning and sometimes a round of preference training after it. The model answers because that is what it learned to do, while the direction is still sitting in the weights.
Because training touches everything, the whole output distribution moves: tone, length, persona, how tightly the model sticks to a system prompt. You pay for a training run and get a new voice. That is what the roleplay tunes on this catalog are, and some of them were trained onto an already abliterated base (82 ids are tagged roleplay as of 2026-11-16). Giving a character a voice takes a tune; projecting out a direction cannot do it.
One of these on the list is Nous Research’s Hermes 4 405B. Its model card, read October 25, 2026, puts the post-training corpus at roughly five million samples, applied to Llama-3.1-405B, and claims lower refusal rates. The output rate right now is n/a per million tokens, taken from the catalog when this page rendered.
A -uncensored suffix describes the outcome, not the method. It might be a fine-tune, a merge of several, or an abliteration the author never labeled. The id says the build answers. Whether it still answers the way it did yesterday is what the daily real-model check records.
Heretic and derestricted builds: gentler abliteration
Two newer words on ids are both abliteration tuned to do less collateral damage, and both get the abliterated tag here.
Heretic, an open-source tool, runs the full procedure with no configuration. It computes one refusal direction per layer and hands the ablation parameters to a TPE optimizer (Optuna), which scores every candidate on two measures together: how often harmful prompts are refused, and the KL divergence on harmless prompts against the original. Because its direction index is a float, it can blend the directions of two neighboring layers rather than committing to one. On this catalog, the builds made with it carry the tool’s name in the id.
Derestricted marks a build made with norm-preserving biprojected abliteration, a variant credited to Jim Lai (grimjim). It strips the refusal component out of each weight row’s direction only, and then restores the row’s original magnitude. The Hugging Face card for GLM-4.6-Derestricted-v3, published by Arli AI and read October 25, 2026, gives the goal as dropping refusals while holding on to GLM-4.6’s reasoning, and attributes the capability loss of plain abliteration to the way it alters neuron magnitudes.
15 builds of this kind are on the catalog today. The model name links to its page, and prices are live.
| Model | Method the id names | Input $/M | Output $/M |
|---|---|---|---|
| DeepSeek V4 Pro Qwen3.5 9B Uncensored Heretic | Heretic, automatic abliteration | $1.12 | $1.12 |
| DeepSeek V4.1 Flash 1017 Heretic | Heretic, automatic abliteration | $0.825 | $3.30 |
| DeepSeek V4.1 Flash 1017 Heretic Thinking | Heretic, automatic abliteration | $0.825 | $3.30 |
| Gemma 31B Heretic Uncensored | Heretic, automatic abliteration | $1.50 | $8.25 |
| Gemma 4 26B-A4B ChimeraX Heretic | Heretic, automatic abliteration | $0.484 | $1.52 |
| Gemma 4 31B Gembrain Heretic | Heretic, automatic abliteration | $1.224 | $1.224 |
| Gemma 4 31B GemBrain Heretic | Heretic, automatic abliteration | $0.40 | $1.76 |
| Gemma 4 31B SDFT Heretic RP | Heretic, automatic abliteration | $1.224 | $1.224 |
| GLM-4.6 Derestricted v5 | Norm-preserving biprojected abliteration | $1.40 | $5.25 |
| Qwen3.5 27B Blossom V6.4 Derestricted | Norm-preserving biprojected abliteration | $1.071 | $1.071 |
| Qwen3.5 27B BlueStar v3 Derestricted | Norm-preserving biprojected abliteration | $1.071 | $1.071 |
| Qwen3.5 27B Claude 4.6 Opus Reasoning Distilled Derestricted | Norm-preserving biprojected abliteration | $1.0389 | $1.0389 |
| Qwen3.5 27B Claude 4.6 Opus Reasoning Distilled Derestricted Lite | Norm-preserving biprojected abliteration | $1.071 | $1.071 |
| Qwen3.5 27B Queen Derestricted | Norm-preserving biprojected abliteration | $1.071 | $1.071 |
| Qwen3.8 27B Heretic | Heretic, automatic abliteration | $1.25 | $3.75 |
What the live catalog shows today
Every count here is read from the catalog tags when the page renders, on 2026-11-16. 170 uncensored builds are listed. 38 of them carry the abliterated tag, split into 23 with abliterated or obliterated in the id and 15 Heretic, derestricted and unshackled builds from the table above. Another 37 are uncensored builds tagged neither roleplay nor security; the roleplay tunes number 82, the security builds 7.
Once a day, one fixed prompt goes to every model and the answer is compared with the last one. The November 16, 2026 run at 02:56 UTC reached 169 of these builds: 147 still answered the same way, 19 had changed, and 3 have not been checked yet. The proof page has every verdict and a CSV.
Price is set per build, so two edits of one base can cost very different amounts. Take 20M input tokens and 5M output tokens in a month, about what a busy chat client or a single coding agent uses: GLM-5.3 Flash Abliterated comes to $19.01, GLM-5.3 Abliterated to $65.48 and GLM-5.3 Uncensored to $9.87, each multiplied from the per-million rates the tables below showed on 2026-11-16. Decide on the edit first, then compare the rows within it.
Ids that say abliterated
Ids that say uncensored, method unstated
The full model list adds the roleplay tunes and the security builds, each with its context window and wire format.
Which one to run
Let the edit decide. Abliteration keeps the base and removes the refusals; a tune swaps in a new voice. Most readers fall into one of three cases.
Chat and roleplay
In SillyTavern, Janitor AI or another persona client, a roleplay tune wins when character and pacing count for more than the base model’s knowledge. An abliterated build wins when you would rather keep the stock voice and range and only lose the refusals; try DeepSeek V4.1 Flash Abliterated, or GLM-5.3 Flash Apex Abliterated if you need its 1,048,576-token context. Look at the context column before a long session: abliterated builds here span 8,192 to 1,048,576 tokens. 4 builds have had both edits, a roleplay tune trained onto an abliterated base, Gemma 4 26B-A4B ChimeraX Heretic among them. No client at all? /chat runs these builds in the browser with three free turns a day and no account, and the roleplay page covers characters. The uncensored LLM leaderboard orders all the builds by the current check and price, and keeps a separate best abliterated LLM list.
Coding
Pick an abliterated build of a base that already codes well. A tune shifts the output distribution, which in code means new habits and new defaults; an abliteration moves one direction and leaves the base model’s coding behavior mostly where it was. Today that means GLM-5.3 Flash Abliterated, GLM-5.3 Abliterated and MiMo V2.6 Flash Abliterated. GLM-5.3 Flash Code Uncensored is the single coding build whose id names the result rather than the method. Set the base URL in OpenCode or Cline as the docs show, then enter the id.
Security work
The security builds are tuned for offensive-security and malware-analysis questions, within the acceptable-use terms. When you want one general model that can also take red-team work, an abliterated build on a current base handles it with no jailbreak prompt to keep alive; AI with no restrictions goes into that case.
| Cyber Frontier | Security | $5.25 | $15.75 |
| Gemma 4 26B A4B Cybersecurity | Security | $0.4224 | $1.3376 |
| GLM-5.3 Flash Cybersecurity | Security | $0.5093 | $1.6975 |
| OrcaCyber MiniMax M3.1 Flash | Security | $2.00 | $4.00 |
| OrcaCyber Qwen3.8 Flash Next | Security | $2.00 | $4.00 |
| OrcaCyber Zero 1.5 | Security | $7.50 | $18.75 |
| Qwen3.8 27B Cybersecurity | Security | $0.40 | $2.40 |
Compare all three with one loop
One OpenAI-format endpoint and one key serve every build, so a test across the three kinds is a loop over ids. Read each answer for the two things an edit can change: does the model begin with the answer or with a refusal, and how far has the voice drifted from the stock model. An abliteration should sound like its base; a tune should sound like the dataset it was trained on. The settled charge rides on each response in a header described in the docs, so the same loop prices the three builds on your prompt. The playground does the same comparison without code.
from openai import OpenAI client = OpenAI(base_url="https://api.unrestrictable.ai/v1", api_key="sk-unr-live-…") # One prompt, three kinds of build; swap the id and nothing else.for model in ["deepseek-v4.1-flash-abliterated", "glm-5.3-uncensored", "glm-5.3-flash-cybersecurity"]: reply = client.chat.completions.create( model=model, messages=[{"role": "user", "content": "In four sentences, what does directional ablation change in a transformer?"}], ) print(model, "->", reply.choices[0].message.content[:160])Reading an id
- abliterated, obliterated: one edit, two spellings. The refusal direction has been projected out of the weights.
- heretic, derestricted: abliteration with the damage held down. Heretic optimizes each run against refusal count and KL divergence; derestricted restores the weight magnitudes after the projection.
- uncensored: the outcome, not the method. Look for the method on the model card if the author wrote one; otherwise the daily check is what you have.
- a roleplay name on an abliterated base: two edits in sequence. First the refusals came out, then the tune put a voice in.
- jailbroken: not a kind of build. A jailbreak is a prompt that talks a stock model around its refusal direction, and it breaks as soon as the prompt changes. No id on this page needs one.
Builds leave the catalog from time to time, and the tables here drop an id the day it goes. A build whose answers change shows as “changed” on the proof page the following day, and each model page carries its own verdict.
Questions
They do not. Abliterated names a method: one direction projected out of the weights, with the rest of the base model untouched. Uncensored names an outcome, most often a fine-tune retrained on answers with the refusals removed, which also changes the voice.
Abliterated, on a base that codes well. Because the edit removes one direction, the base model’s coding style and defaults survive; a fine-tune carries them off toward its training data. See the OpenCode and Cline settings.
Take a roleplay tune if the character’s voice and pacing come first, since those are what it was trained for. An abliterated build if you want the base model’s knowledge with the refusals gone. A few builds on the catalog are both at once, a roleplay tune trained onto an abliterated base. See the uncensored LLM leaderboard.
A little, and it is measurable. On public Gemma 3 12B abliterations the Heretic README reports KL divergence from the original of 0.45 and 1.04 on harmless prompts, and 0.16 for Heretic’s own build, with each one still refusing 3 of 100 harmful prompts. A fine-tune changes the voice by design, so no single drift number exists for it.
Abliterated. Heretic automates the abliteration and searches its parameters against refusal count and KL divergence, so the refusal is projected out of the weights rather than trained away. The catalog tags its builds abliterated.
No. Abliterated is the term coined for the weight edit; obliterated is the everyday English word. Authors use either for the same method, and the catalog tags both abliterated.
Every model above is on the uncensored model list with its live price per million tokens. Per-token pricing sets the whole catalog side by side, and the family directory groups the builds by base model.