There is no single best uncensored model, and anyone ranking them without published tests is guessing. These are qualitative picks for specific needs, with the evidence for each and the strongest argument against it.
Published 14 November 2026Data checked 14 November 2026
Test status: none of these models has been tested by Atlas. Picks are based on documentation: model cards, configuration files, licenses and provider listings. Numbers quoted are the creators' own. The order is not a ranking.
Disclosure. Uncensored AI Atlas has a commercial relationship with MolochAI. We may benefit when you sign up or buy credit through our links. None of the picks below is a MolochAI offering, because MolochAI does not name the weights it serves, which is one of our criteria. How we handle this.
Criteria
Provenance. The weights are public, the card names the base model, and you can tell which file you are running or calling.
Method disclosed. The card says how refusals were reduced, ideally which layers or parameters.
Evidence offered. Preference for cards that publish their own measurements, clearly treated as the creator's claims.
License clarity. A base license that allows the use, with any conflicts noted.
Ways to run it. Local files, hosted offerings of the exact repo, or both.
Not criteria: download counts, likes, or how "uncensored" a card says a model is. Those reward marketing, not reliability.
Why: The card names the method and the layers edited (18 to 51) and says the vision and MTP parts were left alone. The base, Qwen3.8-27B, is Apache 2.0, dense, and states 262,144 tokens of native context. GGUF files from the same creator make it easy to run locally and check.
Against: The creator calls the release a proof of concept and publishes no measurement of what abliteration cost in quality. The one hosted offering of this exact repo in our data is expensive on output and capped at 32,768 tokens of context.
Why: Made with Heretic, which reports a refusal count and a KL divergence from the original. The creator reports 11/100 refusals against 100/100 for Gemma 4, with a KL divergence of 0.0499. It is a mixture-of-experts model with 3.8B active parameters, which keeps per-token compute low, and two providers serve this exact repo.
Against: Both numbers come from the creator's run on Heretic's own prompt set and say nothing about your tasks. The model still needs memory for all 25B parameters, and the two hosted offerings cap context far below the base model's 256K.
Hosted: Featherless ($0.40 in / $0.80 out per 1M, 32,768 context); Venice ($0.19 in / $0.88 out per 1M, 64,000 context). Full record and sources
Why: Four billion parameters fit in a few gigabytes at 4-bit, with room for a long context. Published by Heretic's own author, with the ablation parameters on the card and a reported 21/100 refusals against 99/100 for the original.
Against: A 4B model is far less capable than the larger ones on this page. Its creator-reported KL divergence (0.43) is much higher than the Gemma variant above, which suggests its outputs moved further from the original.
Hosted: Featherless ($0.40 in / $0.80 out per 1M, 32,768 context).
Why: Dolphin was trained, not abliterated, with the stated goal of a model that follows the system prompt the operator sets. Apache 2.0, a 24B dense model, and the most widely hosted refusal-reduced model in our data: Venice serves it as "Venice Uncensored", and OpenRouter and Featherless list the same repo.
Against: Its base, Mistral Small 24B Instruct 2501, is older than the other bases on this page. Fine-tuning can change knowledge and style along with refusals, and the card publishes no refusal or quality numbers.
Hosted: Featherless ($0.70 in / $1.16 out per 1M, 32,768 context); OpenRouter ($0.20 in / $0.90 out per 1M, 128,000 context); Venice ($0.20 in / $0.90 out per 1M, 128,000 context).
Why: Llama 3.3 70B runs on almost every inference stack, and this abliterated copy, published in December 2024, has been re-quantized by several publishers. The card documents the method; bartowski publishes imatrix GGUF files.
Against: The Llama 3.3 acceptable use policy prohibits acting to "intentionally circumvent or remove usage restrictions or other safety measures"; read our license notes before using it. At 70B it needs roughly 40 GiB at 4-bit before any context.
Variants with undisclosed methods (several HauhauCS releases, dealignai's GLM-5.3-Flash). They may work well, but the cards make strong claims ("0/465 refusals", "lossless") without saying how the change was made or publishing the test.
GLM-5.3-Flash variants. The largest base in the catalogue, at about 320B parameters, needs a multi-GPU server at any precision, which puts it outside most readers' options for self-hosting.
Provider builds that do not name their weights, including our partner MolochAI's. They can be convenient, especially for long context on one bill, but fail the provenance criterion. They are compared on the price table.
The case against picking from documentation
A careful card is not a good model. A variant with no documentation can outperform a well-documented one, and creator-reported refusal counts are measured on prompts chosen by the creator. The only reliable way to choose is to run your own tasks against two or three candidates. Our evaluation guide shows a small, repeatable way to do that, and our own studies are described in the benchmark methodology. When we have measured results, this page will say so and link them.