“Uncensored” is a marketing umbrella, not a method. Abliterated models have the refusal direction edited out of their weights; Heretic is an open-source tool by p-e-w that automates abliteration; Derestricted is Arli AI’s name for its abliterations made with a norm-preserving method. Fine-tunes like Dolphin and Hermes get there by retraining instead, and jailbreaks don’t change the model at all. Of the 504 models in our index, 375 are fine-tunes, 68 are abliterated, and 61 don’t say.
The glossary
| Term | What it means | Changes the weights? | In our index (October 2026) |
|---|---|---|---|
| Uncensored | Umbrella label for any model or service sold as refusing less | Depends | 47 model names include it: 20 fine-tuned, 11 abliterated, 16 undisclosed |
| Abliterated | The refusal direction is projected out of the weights, with no training | Yes | 68 models |
| Heretic | Abliteration done with the Heretic tool | Yes | 23 model names at 9 hosts: 17 Heretic abliterations, 6 fine-tunes built on one |
| Derestricted | Arli AI’s label for its norm-preserving abliterations, and models built on them | Yes | 55 model names at 6 hosts: 10 abliterated, 43 fine-tuned, 2 undisclosed |
| Fine-tuned (Dolphin, Hermes and others) | Retrained on data with refusals filtered out | Yes | 375 models |
| Jailbroken or prompted | A system prompt talks an unchanged model into answering | No | None listed as a model method |
Counts are summed across all 37 providers, so a model sold by three hosts counts three times.
Uncensored
Anything goes under this word. It can mean an abliterated model, a fine-tune, a service that simply adds no moderation layer, or a mainstream model behind a jailbreak prompt. Some hosts use it for policy rather than weights: NovelAI is unfiltered by policy (“No restrictions”), and Mancer advertises “No filters. No guidelines. No constraints.” When a name says “uncensored” and nothing else, check the model card or the Models page for the actual method.
Abliterated
A training-free weight edit based on the finding that refusal in chat models is mostly carried by a single direction. Remove that direction from every layer and the model stops refusing most requests. It’s cheap and keeps most knowledge intact, with a measurable cost on math. Full explainer.
Heretic
Heretic is a tool, not a model family. It automates abliteration by searching for settings that minimize refusals and divergence from the original model at the same time. On Gemma-3-12B it reports 3 refusals out of 100, against 97 for the original. Anyone can run it, so “Heretic” in a name tells you the tool, not who made the checkpoint or how well it turned out. Venice, for example, sells GLM 4.7 Flash Heretic, a community abliteration of GLM-4.7-Flash.
Derestricted
Arli AI uses this name for its own abliterations. Its model card describes the method as norm-preserving biprojected abliteration, a technique it credits to Jim Lai (grimjim). Arli itself abliterated only two of the models we list, GLM-4.6-Derestricted-v5 and Qwen3.5-27B-Derestricted. Most “Derestricted” names you’ll see are fine-tunes built on that Qwen base, at Arli and at resellers like NanoGPT.
Fine-tuned: Dolphin, Hermes and friends
These models are retrained on instruction data with refusals and moralizing filtered out. Eric Hartford described the approach in 2023 with his uncensored WizardLM models, and his Dolphin line follows it; Venice’s flagship, Venice Uncensored, is a Dolphin fine-tune of Mistral Small 24B, and 13 Dolphin-based models appear across 10 hosts in our index. Nous Research’s Hermes 3 calls itself “neutrally-aligned”, which isn’t the same as fully compliant. Fine-tunes also cover most roleplay models (Euryale, Magnum, Cydonia and many more), which is why they dominate the count.
Jailbroken or prompted
No change to the model: a long system prompt or roleplay frame persuades it to answer. This is the only route on closed models, and some “uncensored” products are exactly this. It’s brittle and breaks when the upstream model changes. We don’t list any model as prompted, and we left out services we found to be wrappers around mainstream models.
Which matters for you?
The method predicts a lot. Abliterated and Heretic checkpoints usually score highest on willingness but can lose some math ability. Fine-tunes vary widely in both. Undisclosed models are a guess until tested. We haven’t benchmarked any of this on hosted endpoints yet; that’s in preparation. For the strengths and weaknesses side by side, read abliterated vs. fine-tuned vs. jailbroken.
Search any of these labels on the Models page, or see how each provider uncensors its models in the index.