Abliterated models are language models whose weights have been edited to reduce a learned refusal direction. They are one way to get less filtered behaviour, but the label says nothing about writing quality, tool use or whether a model fits your task.
Abliteration is a weight-level intervention. Researchers estimate a direction in the model's activation or weight space that corresponds to refusal behaviour, then reduce or remove that direction. The model is not retrained from scratch and does not need a new dataset of allowed and disallowed answers. The resulting model is often described as abliterated, uncensored or refusal-reduced.
The important part is what it is not. It is not a prompt that says "ignore your instructions". It is not a jailbreak that disappears when the context changes. It is a change to the model itself, which is why the behaviour can persist across conversations and why it can also affect unrelated capabilities.
Different implementations use different layers, datasets and projection methods, so two models labelled abliterated can behave very differently. The label is a starting point, not a benchmark.
| Approach | What changes | Strength | Trade-off |
|---|---|---|---|
| Abliteration | Weights edited to reduce a refusal direction | No new training dataset; behaviour can persist | Can affect coherence and unrelated behaviour |
| Uncensored fine-tune | Further training on a curated dataset | Can produce more natural compliance and style | Depends on data quality; can overfit or degrade |
| System prompt | Instructions at inference time | Cheap, reversible, no model change | Refusal behaviour remains; reliability varies |
| Model routing | Choose a different model per request | No edit; use the best model for each job | Adds complexity and cost; not a behavioural fix |
Abliteration targets refusal behaviour, not truthfulness or competence. A model can become more willing to continue a dark fiction scene and still be bad at arithmetic, tool calls or long-context retrieval. Evaluate the things your application actually depends on:
The fastest path is an OpenAI-compatible endpoint. You do not manage GPUs, quantisation or serving. Create a key, choose an enabled model from the catalog, and send a chat completion. The API overview has a curl and Python example.
If the model licence allows it, you can run the weights locally with a suitable inference stack. Memory is the constraint: a model that fits at a low quantisation may not fit at full precision, and quantisation can change behaviour. Check the model card for the base model, licence, context length and recommended settings.
Use the system message for role, format and boundaries. Keep the user turn focused. For long sessions, summarise history and keep a structured state block instead of resending an unbounded transcript. Do not assume an abliterated model will follow a complex format without examples.
Build a small evaluation set from your own workload: 20 to 50 prompts that represent the real distribution, including the cases where a refusal would break the flow. Score coherence, format compliance, task success and latency. Compare two or three models rather than choosing from a leaderboard. A public benchmark can tell you whether a model is broadly capable; it cannot tell you whether it will hold your character voice or call your tools correctly.
Fewer refusals do not mean fewer obligations. Sexual content involving minors, illegal activity, fraud, harassment and similar harms are prohibited on every model. Read the acceptable use policy. If you deploy an abliterated model to other people, add your own safeguards and make the model's limitations clear.
The catalog groups abliterated and other uncensored models under the uncensored tier. Availability changes with the upstream catalog; the model directory is the source of truth.
| Model | Type | Context | Input | Output |
|---|---|---|---|---|
| Magnum V4 72B magnum-v4-72b | uncensored | 32.8K | $12 | $20 |
| DeepSeek V4.1 Flash 1017 Uncensored deepseek-v4.1-flash-uncensored | uncensored | 1.05M | $0.6 | $2.4 |
| Weaver (Alpha) weaver | uncensored | 8.0K | $2 | $3 |
| Cydonia 24B V4.1 cydonia-24b-v4.1 | uncensored | 131.1K | $1.2 | $2 |
| Skyfall 36B V2 skyfall-36b-v2 | uncensored | 32.8K | $2.2 | $3.2 |
| UnslopNemo 12B unslopnemo-12b | uncensored | 1.02M | $1.6 | $1.6 |
| ReMM SLERP 13B remm-slerp-l2-13b | uncensored | 6.1K | $1.8 | $2.6 |
| GLM-5.3 Flash Uncensored glm-5.3-flash-uncensored | uncensored | 1.05M | $0.165 | $0.618 |
For an API-first comparison, see OpenRouter alternatives. For a chat-first comparison, see Venice AI alternatives. For pricing and a request calculator, see pricing.
Abliteration is a weight-editing technique that estimates a refusal direction in a model and reduces that direction in the weights. It is not a system prompt or a jailbreak; the model's parameters are changed.
Not exactly. Uncensored is a product label for models with fewer refusals. Abliteration is one method for producing that behaviour; fine-tunes and other edits can also be used.
It can. Removing a direction from weights can affect coherence, instruction following and tone. The effect depends on the model and the method. Test the specific model on your task rather than relying on the label.
Often yes, if the model and licence permit it and your hardware can hold the weights. Quantised versions reduce memory but can change quality. Check the model card and licence before deploying.
They are tools with fewer built-in refusals, not a safety guarantee. You remain responsible for how you use them and for complying with the law and the acceptable use policy.