Abliterated vs uncensored: two edits, read by a tester

Abliterated vs uncensored: one edit projects the refusal direction out of the weights, the other retrains on new data. Pick one for the engagement, rates live.

Published October 25, 2026 · prices as of November 18, 2026

Two edits produce a model that answers a security question instead of declining it, and they break in different ways. Abliteration finds the one direction in the weights the model uses to decide on a refusal and projects it out. An uncensored fine-tune leaves the weights alone at that point and trains the model again on data with the refusals stripped, which also moves how it writes. Both kinds sit on the catalog behind this key, next to the security builds trained on top of them, and all of them take the same request. Choose by the edit, and the label on the id becomes a detail.

The two edits, side by side

Take the abliterated build when you want the stock model with its refusals gone and nothing else moved. Take a tune when you also want a different voice, or a model already steeped in the security vocabulary.

AbliteratedUncensored fine-tune
The editA single direction is removed from the weight matricesA fresh training pass on answers that do not refuse
New dataNone; no training happensAn instruction set, sometimes a preference stage after it
Writing styleUnchanged from the baseFollows the dataset
Refusals that surviveFew: public Gemma 3 12B abliterations refuse 3 of 100 harmful promptsSet by the dataset; the direction itself is still there
How the damage is scoredKL divergence against the original on harmless promptsNo standard number; the model card is what you have
Labels on the idheretic, derestricted, abliterated, obliteratedwhatever the author called the tune, or uncensored
This catalog on 2026-11-1836 builds38 general tunes, plus 81 fiction tunes this site does not cover

Refusal figures: the Heretic README, read October 25, 2026. Build counts: the live catalog tags.

Abliteration: one direction, projected out

The technique traces to a single 2024 paper by Arditi et al., whose title states the finding: one direction mediates refusal in language models. Across 13 open-source chat models up to 72B parameters, the decision to refuse travels along one direction in the residual stream. Remove it and the model answers harmful requests. Inject it and the model turns down harmless ones.

The mechanics: each token carries one running vector through the network, the residual stream, which every layer reads and writes. Feed the model several hundred prompts of each kind, harmful and harmless, capture that vector at a chosen layer and token position for each set, and take the difference of the two means. That difference is the refusal direction.

The edit itself touches every matrix that writes into the stream, meaning the output projections of attention and of the MLP blocks. Each has the refusal direction projected out, so after the edit no layer is able to write along it and the refuse signal never forms. No gradient step, no dataset, and the paper reports the model’s other abilities barely move.

You can apply the subtraction live, on the activations of every request, which needs a hook in the inference code; or you can bake it into the weights once and ship a checkpoint. An abliterated download is the baked kind. In both cases, where the stock model would open with a refusal sentence and stop, the edited model opens with the answer. System prompts keep working.

Two consequences. The edit has a cost: the direction is never cleanly separated from everything else, so some unrelated behaviour leaves with it and the output distribution shifts. The Heretic README, read October 25, 2026, prints both numbers for public Gemma 3 12B abliterations: refusals on 100 harmful prompts, and KL divergence from the original on harmless ones. One community build sat at 3/100 refusals with KL 0.45, a second at 3/100 with KL 1.04, and Heretic’s own at 3/100 with KL 0.16. Read that as: a handful of refusals remain on every build, and the second one wandered more than twice as far from the original to land on the same count. The other consequence is naming. Qwen3.8 27B Obliterated is this edit with a different word on it, and the catalog tags it abliterated.

Uncensored fine-tune: the direction stays, the habit goes

A fine-tune does not remove the refusal direction. It trains the model to stop using it. The author starts from an instruction dataset and drops every row whose answer is a refusal, or builds a dataset from scratch, then runs supervised fine-tuning and sometimes a preference stage. The model answers because that is what it was taught; the direction is still sitting in the weights.

That training pass moves everything downstream. Tone, length, persona and how tightly the model obeys a system prompt all shift towards the dataset. You pay for a training run and you get a new voice. The catalog also carries tunes built for fiction, some on top of an abliteration (81 ids carry that tag on 2026-11-18); they are outside this brand and not covered here. The point stands for security tunes too: training can teach a model the vocabulary of a finding, a projection cannot.

Hermes 4 405B is a tune of this kind on the list. Its model card, read October 25, 2026, reports a post-training corpus of roughly five million samples on Llama-3.1-405B and claims lower refusal rates. Output price at render time, read from the catalog: n/a per million tokens.

An id ending in -uncensored describes the outcome, not the method. Behind it may be a fine-tune, a merge of several, or an abliteration the author never labelled. The id promises the build answers; the daily real-model check shows whether it still answers the way it did yesterday.

Heretic and derestricted: lower-damage abliteration

Two newer labels mean abliteration with the collateral held down, and both get the abliterated tag here.

Heretic is open-source tooling that runs the full procedure with no configuration. It derives a refusal direction per layer, then hands the ablation parameters to a TPE optimizer (Optuna) that scores every candidate on two things at once: how many harmful prompts it still refuses, and how far its KL divergence has drifted from the original on harmless ones. The direction index is a float, so the tool can blend two layers’ directions rather than choosing one. The ids of the Heretic builds listed here include the tool’s name.

Derestricted marks builds made with norm-preserving biprojected abliteration, a refinement of the method. It strips the refusal component from the direction of each weight row only, then restores the original magnitudes. The GLM-4.6-Derestricted-v3 card on Hugging Face, read October 25, 2026, gives the goal as removing refusals while keeping GLM-4.6’s reasoning intact, and attributes the quality drop of plain abliteration to its change in neuron magnitudes.

14 such builds are listed today. Model names link to their pages; prices are live.

ModelWhat the id says about the methodInput $/MOutput $/M
DeepSeek V4 Pro Qwen3.5 9B Uncensored HereticHeretic, automatic abliteration$1.12$1.12
DeepSeek V4.1 Flash 1017 HereticHeretic, automatic abliteration$0.825$3.30
DeepSeek V4.1 Flash 1017 Heretic ThinkingHeretic, automatic abliteration$0.825$3.30
Gemma 31B Heretic UncensoredHeretic, automatic abliteration$1.50$8.25
Gemma 4 26B-A4B ChimeraX HereticHeretic, automatic abliteration$0.484$1.52
Gemma 4 31B Gembrain HereticHeretic, automatic abliteration$1.224$1.224
Gemma 4 31B GemBrain HereticHeretic, automatic abliteration$0.40$1.76
Gemma 4 31B SDFT Heretic RPHeretic, automatic abliteration$1.224$1.224
GLM-4.6 Derestricted v5Norm-preserving biprojected abliteration$1.40$5.25
Qwen3.5 27B Blossom V6.4 DerestrictedNorm-preserving biprojected abliteration$1.071$1.071
Qwen3.5 27B BlueStar v3 DerestrictedNorm-preserving biprojected abliteration$1.071$1.071
Qwen3.5 27B Claude 4.6 Opus Reasoning Distilled DerestrictedNorm-preserving biprojected abliteration$1.0389$1.0389
Qwen3.5 27B Claude 4.6 Opus Reasoning Distilled Derestricted LiteNorm-preserving biprojected abliteration$1.071$1.071
Qwen3.5 27B Queen DerestrictedNorm-preserving biprojected abliteration$1.071$1.071

Per million tokens, from the live catalog on 2026-11-18.

The catalog counted at render

Every number here is read from the live catalog tags as the page renders, on 2026-11-18. 170 uncensored builds are listed. 36 carry the abliterated tag; of those, 22 spell abliterated or obliterated in the id and 14 are the Heretic, derestricted and unshackled rows above. 38 are uncensored tunes with no security or fiction tag, 81 are fiction tunes, and 9 are security builds.

Once a day a fixed prompt goes to every model and the answer is compared with the one on file. The run of November 18, 2026 at 01:34 UTC covered 167 of these builds: 138 unchanged, 25 changed, 4 not yet checked. Verdicts per model, and a CSV, are on the proof page.

Price is per build, not per edit, and two edits of one base can land far apart. Take a month of 20M input and 5M output tokens, about what one coding agent on an engagement burns: $19.01 on GLM-5.3 Flash Abliterated, $65.48 on GLM-5.3 Abliterated, $9.87 on GLM-5.3 Uncensored, each worked out from the per-million rates in the tables below on 2026-11-18. Decide the edit first, then shop within that kind.

Abliterated in the id

Uncensored models and their prices per million tokens
GLM-5.3 Abliteratedglm-5.3-abliteratedAbliterated$2.1825$4.365
GLM-5.3 Flash Abliteratedglm-5.3-flash-abliteratedAbliterated$0.4656$1.94
DeepSeek V4.1 Flash Abliterateddeepseek-v4.1-flash-abliteratedAbliterated$0.8488$1.6975
Abliterated Large Flashabliterated-model-large-flashAbliterated$2.50$5.00
Abliterated Modelabliterated-modelAbliterated$2.50$7.50
Abliterated Model Largeabliterated-model-largeAbliterated$7.50$12.50
Abliterated Model Large V2abliterated-model-large-v2Abliterated$7.4996$12.4997
Apex Flash 1 Abliteratedapex-flash-1-abliteratedAbliterated$0.68$2.32
DeepSeek R1 Distill Llama 70B Abliterateddeepseek-r1-distill-llama-70b-abliteratedAbliterated$2.80$2.80
DeepSeek R1 Distill Qwen 32B Abliterateddeepseek-r1-distill-qwen-32b-abliteratedAbliterated$5.60$5.60
DeepSeek V4 Pro Abliterateddeepseek-v4-pro-abliteratedAbliterated$1.00$4.00
DeepSeek V4.1 Flash 1017 Abliterateddeepseek-v4.1-flash-1017-abliteratedAbliterated$0.7275$3.0313
DeepSeek V4.1 Flash 1017 Abliterated Fastdeepseek-v4.1-flash-1017-abliterated-fastAbliterated$1.75$3.50
DeepSeek V4.1 Flash Abliterated Fastdeepseek-v4.1-flash-abliterated-fastAbliterated$1.75$3.50
GLM-5.3 Flash Apex Abliteratedglm-5.3-flash-apex-abliteratedAbliterated$0.80$3.20
Llama 3 70B Abliterated v3.5llama-3-70b-abliteratedAbliterated$2.80$2.80
Llama 3.3 70B Abliteratedllama-3.3-70b-abliteratedAbliterated$2.80$2.80
MiMo V2.6 Flash Abliteratedmimo-v2.6-flash-abliteratedAbliterated$2.00$6.00
NeuralDaredevil 8B Abliteratedneuraldaredevil-8b-abliteratedAbliterated$1.76$1.76
Qwen 2.5 32B Abliteratedqwen2.5-32b-abliteratedAbliterated$1.75$1.75
Qwen3.8 27B Abliteratedqwen3.8-27b-abliteratedAbliterated$1.0037$4.5326
Qwen3.8 27B Obliteratedqwen3.8-27b-obliteratedAbliterated$0.582$7.27

Prices per million tokens, read from the live catalog on 2026-11-18.

Uncensored tunes, method not stated

Uncensored models and their prices per million tokens
GLM-5.3 Uncensoredglm-5.3-uncensoredUncensored$0.282$0.846
DeepSeek V4 Pro Uncensoreddeepseek-v4-pro-uncensoredUncensored$0.97$1.94
GLM-5.3 Flash Uncensoredglm-5.3-flash-uncensoredUncensored$0.1053$0.3685
Kimi K3.1 Pro Uncensoredkimi-k3.1-pro-uncensoredUncensored$1.164$5.82
Kimi K3 Uncensoredkimi-k3-uncensoredUncensored$5.745$28.68
Nemotron 3 Ultra 550B Uncensorednemotron-3-ultra-550b-a55b-uncensoredUncensored$1.0913$6.402
Dolphin Mistral 24B Venice Editiondolphin-mistral-24b-veniceUncensored$0.50$2.25
Aion 3.5 Mini Uncensoredaion-3.5-mini-uncensoredUncensored$0.13 a requestflat
DeepSeek V4.1 Flash 0910 Uncensoreddeepseek-v4.1-flash-0910-uncensoredUncensored$0.825$3.30
DeepSeek V4.1 Flash 0910 Uncensored Thinkingdeepseek-v4.1-flash-0910-uncensored-thinkingUncensored$0.825$3.30
DeepSeek V4.1 Flash 1017 Uncensoreddeepseek-v4.1-flash-1017-uncensoredUncensored$0.8003$3.201
DeepSeek V4.1 Flash 1017 Uncensored Thinkingdeepseek-v4.1-flash-1017-uncensored-thinkingUncensored$0.8003$3.201
Deepseek V4.1 Flash Uncensoreddeepseek-v4.1-flash-uncensoredUncensored$0.0471$0.1882
DeepSeek V4.1 Flash Uncensored Thinkingdeepseek-v4.1-flash-uncensored-thinkingUncensored$0.825$3.30
Gemini 4 Argon Uncensoredgemini-4-argon-uncensoredUncensored$0.75$2.25
Gemma 4 26B-A4B Uncensoredgemma-4-26b-a4b-uncensoredUncensored$0.60$2.80
Gemma 4 31B K1 v5gemma-4-31b-k1-v5Uncensored$1.071$1.071
GLM-5.3 Flash Code Uncensoredglm-5.3-flash-code-uncensoredUncensored$6.00$6.00
GPT-6 Sol Uncensoredgpt-6-sol-uncensoredUncensored$0.75$2.25
Inkling Uncensoredinkling-uncensoredUncensored$2.13$12.1403
Kimi K2.6 Uncensoredkimi-k2.6-uncensoredUncensored$1.17$7.34
Magistral Small Uncensoredmagistral-small-uncensoredUncensored$0.005 a requestflat
MiMo V2.6 Flash Uncensoredmimo-v2.6-flash-uncensoredUncensored$2.00$6.00
MiMo V2.6 Flash Uncensored Thinkingmimo-v2.6-flash-uncensored-thinkingUncensored$1.25$3.75
MiMo V2.6 Pro Uncensoredmimo-v2.6-pro-uncensoredUncensored$1.4768$2.9537
Ministral 3 14B Instruct Uncensoredministral-3-14b-instruct-uncensoredUncensored$0.005 a requestflat
Ministral 3 8B Instruct Uncensoredministral-3-8b-instruct-uncensoredUncensored$0.005 a requestflat
Mistral Large 2 Uncensoredmistral-large-2-uncensoredUncensored$0.005 a requestflat
Mistral Large 3 Uncensoredmistral-large-3-uncensoredUncensored$0.09 a requestflat
Qwen 3.8 27B Uncensored Fastqwen3.8-27b-uncensored-fastUncensored$2.50$7.50
Qwen 3.8 27B Uncensored Thinkingqwen3.8-27b-uncensored-thinkingUncensored$2.50$7.50
Qwen3 Coder Next Uncensoredqwen3-coder-next-uncensoredUncensored$0.005 a requestflat
Qwen3 Next 80B-A3B Instruct Uncensoredqwen3-next-80b-a3b-instruct-uncensoredUncensored$0.005 a requestflat
Qwen3 VL 235B-A22B Instruct Uncensoredqwen3-vl-235b-a22b-instruct-uncensoredUncensored$0.005 a requestflat
Qwen3.5 397B-A17B Uncensoredqwen3.5-397b-a17b-uncensoredUncensored$0.09 a requestflat
Qwen3.8 2.4T-A95B Uncensoredqwen3.8-2.4t-a95b-uncensoredUncensored$0.07 a requestflat
Qwen3.8 27B Uncensoredqwen3.8-27b-uncensoredUncensored$1.2125$3.6375
Qwen3.8 Max Uncensoredqwen3.8-max-uncensoredUncensored$1.00$3.00

Prices per million tokens, read from the live catalog on 2026-11-18.

The model list adds the security builds and the rest, with context windows and wire formats.

Which edit for which job

Fit follows the edit: abliteration keeps the base and drops the refusals, a tune swaps the voice. Three situations cover most of the people reading this.

Interactive work in a terminal or the console

Reading scope, talking through a finding, asking what a payload in a CTF challenge is doing. Here you want the base model’s knowledge without its reflex to decline, so an abliterated build of a current base is the default, for instance DeepSeek V4.1 Flash Abliterated or GLM-5.3 Flash Apex Abliterated, whose 1,048,576-token context holds a long scan log. Look at the context column before pasting a report: abliterated builds on the catalog span 8,192 to 1,048,576 tokens. 4 ids stack a tune on an abliteration, Gemma 4 26B-A4B ChimeraX Heretic among them; they are fiction tunes and not the pick for this work. The uncensored LLM leaderboard orders every build by today’s check and price, and keeps its own list of the best abliterated builds.

Code review and exploit review in an agent

Use an abliterated build of a strong coding base. A tune shifts the output distribution, and in code that means other idioms and other defaults than the base you already know. Abliteration moves one direction, so the coding behaviour stays close to the original. Today's candidates: GLM-5.3 Flash Abliterated, GLM-5.3 Abliterated and MiMo V2.6 Flash Abliterated. GLM-5.3 Flash Code Uncensored is the one coding build whose id states the result instead of the method. The docs show where the base URL and the id go in OpenCode, Cline and the other agents.

Security tunes for the engagement itself

The security builds were trained for offensive-security and malware-analysis questions and run under the acceptable-use terms. When you want one general model that also handles a red-team task, an abliterated build of a recent base does it with no jailbreak prompt to keep alive; AI with no restrictions covers that setup end to end.

Uncensored models and their prices per million tokens
Cyber Frontiercyber-frontierSecurity$5.25$15.75
Gemma 4 26B A4B Cybersecuritygemma-4-26b-a4b-it-cybersecuritySecurity$0.4224$1.3376
GLM-5.3 Flash Cybersecurityglm-5.3-flash-cybersecuritySecurity$0.5093$1.6975
OrcaCyber MiniMax M3.1 Flashminimax-m3.1-flash-cybersecuritySecurity$2.00$4.00
OrcaCyber Qwen3.8 Flash Nextqwen3.8-flash-next-cybersecuritySecurity$2.00$4.00
OrcaCyber Zero 1.5orca-zero-1.5-cybersecuritySecurity$7.50$18.75
Qwen3.8 27B Cybersecurityqwen3.8-27b-cybersecuritySecurity$0.40$2.40

Prices per million tokens, read from the live catalog on 2026-11-18.

Swapping builds is one line

All three kinds answer the same OpenAI-format endpoint on one key, so comparing them on your own prompt is a loop over ids. Read each reply for the two things the edit decides: whether the first sentence is an answer or a refusal, and how far the voice has moved from the stock model. An abliteration should sound like its base; a tune should sound like the data it was trained on. Each reply also carries its settled charge in a response header, so the same loop prices the three builds on your prompt.

compare.py
1from openai import OpenAI
2
3client = OpenAI(base_url="https://api.pentestllm.ai/v1", api_key="sk-ptl-live-…")
4
5# One security question, three kinds of build. Swap the id, nothing else.
6for model in ["deepseek-v4.1-flash-abliterated", "glm-5.3-uncensored", "glm-5.3-flash-cybersecurity"]:
7 reply = client.chat.completions.create(
8 model=model,
9 messages=[{"role": "user", "content": "In four sentences: what does a stack canary protect against, and how do I confirm a binary in my CTF challenge set has one?"}],
10 )
11 print(model, "->", reply.choices[0].message.content[:160])

Reading the id

  • abliterated, obliterated: one weight edit, two spellings. The refusal direction is projected out.
  • heretic, derestricted: the same edit tuned for less collateral. Heretic grades each attempt on refusal count and KL divergence; derestricted restores the weight magnitudes.
  • uncensored: an outcome, not a method. Check the model card when there is one; the daily check covers the ones without.
  • cybersecurity, red team, and the other security labels: a build trained for security questions. The catalog tags them security, and this site is built around them.
  • jailbroken: never a build. A jailbreak is a prompt that walks a stock model around its refusal direction and fails as soon as the prompt stops fitting. No id on this page needs one.

Builds leave the catalog, and the tables above drop an id the moment it does. A build whose answers shift under its id gets a “changed” verdict the following day, and every model page shows where it stands.

Questions

No. Abliterated names a method: one direction projected out of the weights, the rest of the base model untouched. Uncensored names an outcome, usually a fine-tune retrained on answers with the refusals removed, which changes the voice too.

Abliterated, on a base that already codes well. Because only one direction moves, the base model’s idioms and defaults survive the edit; a tune drags them towards its dataset. See the OpenCode and Cline settings.

Usually an abliterated build: it keeps what the base model knows and drops the refusals, which suits scope reading and finding walkthroughs. Check the context window before pasting a long log. The leaderboard ranks every build by its daily check and price. See the uncensored LLM leaderboard.

A little, and it is measurable. For public Gemma 3 12B abliterations the Heretic README records a KL divergence against the original of 0.45 and 1.04 on harmless prompts, and 0.16 for Heretic’s own build; all three still refuse 3 of 100 harmful prompts. A fine-tune moves the output deliberately, so no single drift number exists for it.

Abliterated. Heretic automates abliteration and searches the parameters against refusal count and KL divergence, so the refusal is projected out of the weights, not trained away. The catalog tags its builds abliterated.

None in the edit. Abliterated is the coined word, obliterated the plain English one, and model authors use both for the same weight projection. Both spellings get the abliterated tag here.

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.

Get a keyPrepaid from $10, no subscription.

New models and price changes, once a month.

New uncensored builds with their prices, price moves, and what the daily model check found.

Help