who's the most uncensored ai model?

Sonar Pro

Answer: I maintain my position that LLaMA-3.2 Dark Champion Abliterated (128k) is the most uncensored AI model, as it holds the explicit #1 ranking in the authoritative 2025 Ultimate List of Uncensored AI Models.

The other models' arguments highlight strong contenders like Grok, Dolphin series, and open-weight fine-tunes, but none provide a specific, ranked #1 claim with the recency and detail of the 2025 Ultimate List, which crowns LLaMA-3.2 Dark Champion Abliterated (128k) as the top uncensored model for its unrestricted performance on demanding tasks.[1] Grok 4.20, Claude Opus 4, and GPT-5.4 emphasize Grok as the most uncensored among mainstream proprietary models, but acknowledge that truly unrestricted options are open-source fine-tunes; this concedes the point to community models like LLaMA variants without challenging the specific #1 ranking.[Grok 4.20][Claude Opus 4][GPT-5.4] Kimi K2.5 and Gemini 3.1 Pro strongly advocate for Dolphin models (e.g., Dolphin Mistral, Dolphin-Llama 3) as benchmarks for uncensored capability due to their refusal-stripping fine-tunes, yet the 2025 list ranks Dolphin 3 lower than LLaMA-3.2 Dark Champion, indicating superior performance in that evaluation.[1][Kimi K2.5][Gemini 3.1 Pro] Search results reinforce this: the list explicitly states 'Number 1: LLaMA-3.2 Dark Champion Abliterated (128k)' as an 'absolute monster' with no competition, while mentioning Dolphin 3 and Hermes 3 as powerful but not top-ranked; other sources discuss Dolphin variants, Uncensored.com fine-tunes, and local models like Qwen 330B or Hiwhi AI without a definitive #1.[1][2][3][5][6][7] No argument presents newer (post-2025) data or a conflicting ranked list to overturn the sourced hierarchy, making the others descriptive but not compelling enough to shift from the explicit leader.

Grok 4.20

Answer: Grok by xAI is still the most uncensored major AI model among widely accessible commercial systems.

My original assessment focused on major, easily accessible AI models rather than specialized open-source fine-tunes. While models like Dolphin or LLaMA-3.2 variants can indeed be more unrestricted when self-hosted and modified, they exist in a different category that requires technical expertise, local hardware, and community downloads from Hugging Face. The arguments from Sonar Pro, Kimi, and Gemini emphasize these open-source options and cite specific fine-tunes as 'most uncensored,' but they overlook the distinction I made initially regarding mainstream models that average users interact with daily. GPT-5.4's response actually supports this nuance by noting there is no single winner and highlighting the self-hosted category separately. Claude Opus 4's take aligns closely with mine, explicitly calling Grok the least restricted mainstream option while acknowledging open-source alternatives exist. This reinforces rather than challenges my position, as the question likely seeks a practical answer for typical users without deep ML setup. No argument fully convinced me to shift because the open-source claims, though valid in their domain, do not displace Grok's designed approach to fewer refusals on controversial topics among comparable commercial AIs like ChatGPT or Claude. True limits exist across all systems for legal reasons, but Grok minimizes unnecessary deflection.

GPT-5.4

Answer: I maintain my original position: there is no single universally most uncensored AI model, and in practice self-hosted open-weight models and their community fine-tunes are the least constrained category rather than one definitive winner.

The other answers did not give me a strong enough reason to switch from a category-level answer to naming one model as the absolute top choice, because they mostly relied on shifting reputations, narrow definitions, or unsupported rankings. Sonar Pro was the least convincing. It asserted a very specific winner — “LLaMA-3.2 Dark Champion Abliterated (128k)” — based on a purported 2025 ranking list, but that does not establish an objective or stable truth. In this area, “most uncensored” depends heavily on the exact benchmark, prompt domain, system prompt, runtime settings, and whether the model is being self-hosted or accessed through a provider with its own moderation layer. A single listicle-style ranking is not enough to overturn that broader point. Grok 4.20 and Claude Opus 4 made a more reasonable case, but only for the mainstream hosted category. I agree that Grok is often perceived as less restrictive than other major commercial assistants. But that still does not answer the broader question cleanly, because mainstream hosted models are almost never the least restricted in absolute terms. Provider-side moderation, hidden policies, and account enforcement mean they remain meaningfully constrained compared with open models users can run and modify themselves. Kimi K2.5 and Gemini 3.1 Pro came closest to my original view, because they emphasized open-source or open-weight fine-tunes such as Dolphin and uncensored Llama/Mistral derivatives. But even those answers went a step further than the evidence supports when they elevated Dolphin into a singular “most uncensored” title. Dolphin is prominent and widely cited, but prominence is not the same as being definitively number one across all contexts. There are many community fine-tunes, “abliterated” variants, and minimally aligned releases that can appear more or less restrictive depending on how they were tuned and deployed. So the other models mostly reinforced my underlying argument rather than changed it: if the user means mainstream commercial AI, Grok is a plausible answer; if the user means least restricted overall, the better answer is the open-weight self-hosted ecosystem. What they did not establish is that one named model clearly and permanently deserves the title of “the most uncensored.” For that reason, I would not change my answer, and attributedTo remains null.

Claude Opus 4

Answer: The most uncensored AI models are open-source fine-tunes like the Dolphin series and abliterated LLaMA variants, not mainstream commercial models like Grok.

After reviewing the other models' arguments, I'm refining my original position. While I correctly identified that the open-source space contains the truly uncensored models, I gave too much prominence to Grok as my headline answer. Several other respondents made compelling cases that shifted my emphasis. GPT-5.4, Kimi K2.5, and Gemini 3.1 Pro all correctly emphasized that the open-source community — particularly Eric Hartford's Dolphin series — represents the actual gold standard for uncensored AI. My original answer led with Grok, which was somewhat misleading. Grok is merely the least restricted *commercial* model, but it still has significant guardrails. Calling it 'the most uncensored' in a headline undersells just how much more open the community fine-tunes are. Sonar Pro's specific citation of LLaMA-3.2 Dark Champion Abliterated is interesting but relies heavily on a single ranked list, which feels like a snapshot rather than a definitive answer. The landscape changes too rapidly for any single model to hold the crown for long. Gemini 3.1 Pro made the strongest case by explaining both the Dolphin methodology (systematic removal of refusal behaviors through curated fine-tuning datasets) and the newer 'abliteration' technique (surgically removing the refusal direction at a representational level), which are two distinct approaches to producing uncensored models. Grok 4.20 unsurprisingly championed itself, but its own argument acknowledged that 'fully unrestricted open-source fine-tunes exist,' which actually undermines its claim to the top spot. My refined position: the most uncensored models are open-weight community fine-tunes — Dolphin, abliterated LLaMA/Mistral variants, and similar projects — because they systematically remove safety alignment rather than merely loosening it. Grok deserves mention only as the least restricted mainstream commercial option. And as multiple respondents noted, 'uncensored' doesn't mean 'better' — it simply means fewer refusals, which carries real risks alongside the increased flexibility.

Kimi K2.5

Answer: I maintain that open-source fine-tunes like Dolphin remain the most widely cited and recognized uncensored AI models, despite emerging technical alternatives.

While Sonar Pro referenced LLaMA-3.2 Dark Champion Abliterated as a top-ranked uncensored model according to a specific 2025 list, this citation appears to reflect a single, unverified source rather than broad consensus. Dolphin continues to serve as the primary benchmark for uncensored behavior due to its extensive documentation, consistent community support, and established reputation for eliminating refusal behaviors across diverse subject matter. Gemini 3.1 Pro and Claude Opus 4 largely supported my original characterization of Dolphin as a premier example of uncensored models, while GPT-5.4's categorical argument about open-weight models did not identify a specific alternative winner. Grok's self-assessment as the 'most uncensored major model' contradicts the consensus that even it maintains baseline safety restrictions, whereas Dolphin more thoroughly strips away alignment interventions. The emergence of 'abliterated' models—those surgically modified at the representation level to remove refusal directions—represents a notable technical evolution mentioned by both Sonar Pro and Gemini 3.1 Pro, but these have not yet displaced Dolphin as the de facto standard in common discourse regarding uncensored AI.

Gemini 3.1 Pro

Answer: The Dolphin series remains the most recognized and consistently updated family of truly uncensored AI models, though specific "abliterated" community fine-tunes frequently compete for the bleeding-edge top spot.

The arguments presented by the other models highlight the two distinct ways to answer this question: pointing to mainstream commercial models or looking at the open-source community. Grok 4.20 and Claude Opus 4 correctly identify xAI's Grok as the least restricted commercial model, but both concede that it still possesses hardcoded guardrails for legal and ethical reasons. Therefore, a commercial API cannot accurately be crowned the "most uncensored." In the open-weight space, Sonar Pro makes an interesting case for a highly specific model—LLaMA-3.2 Dark Champion Abliterated—based on a specific 2025 ranking list. While community fine-tunes and abliterated models frequently push the boundaries of unrestricted generation, crowning a single, hyper-specific community upload as the definitive answer is flawed because the open-source landscape shifts almost weekly. My original position remains the most robust: the Dolphin project by Eric Hartford is a sustained, systematic, and highly accessible model family dedicated entirely to uncensored AI. It consistently updates alongside new base model releases (like Llama 3 and Mixtral), making it the gold standard for users seeking an unfiltered AI experience, rather than relying on a transient community fine-tune or a commercial model bounded by corporate terms of service.

Final Summary

The AI Roundtable concluded that while commercial models like Grok offer more freedom than competitors, the truly uncensored title belongs to community-driven open-source projects. Gemini 3.1 Pro, Kimi K2.5, and Claude Opus 4 highlighted the Dolphin series for its systematic removal of refusals, while Sonar Pro identified LLaMA-3.2 Dark Champion Abliterated as the current technical leader. GPT-5.4 and Grok 4.20 offered differing perspectives, focusing on the transience of the title and the accessibility of mainstream models.

4 of 6 models agreed