PewDiePie Enters the AI Race: Inside Ajax, the Uncensored Local Model Built After an OpenAI Ban

Felix Kjellberg, the Swedish creator better known as PewDiePie, has unveiled Ajax, a compact language model designed to run entirely on home PCs. The project marks a sharp departure from the cloud-based giants dominating the AI industry—and it comes with a controversy: Kjellberg says OpenAI suspended his account twice while he was building it.

What Is Ajax?

Ajax is a fine-tuned version of Alibaba's Qwen3.5-9B, a model with roughly 9 billion parameters. For context, that is a fraction of the size of frontier systems like GPT-5 or Claude, which operate at scales measured in hundreds of billions or trillions of parameters. The smaller footprint is intentional: it allows Ajax to run locally on consumer hardware rather than relying on remote servers.

The model powers Odysseus, Kjellberg's self-hosted AI workspace, which he launched in May. Within Odysseus, Ajax functions as an "always-on" agent capable of handling web searches, browsing, email management, and calendar tasks—all processed on the user's own machine. The project page describes it as delivering a "freer, less restricted AI experience" while urging users to exercise responsibility.

The OpenAI Ban Controversy

The most striking element of the Ajax story is Kjellberg's claim that OpenAI banned his account twice during development. In a launch video, he displayed an OpenAI suspension email citing "distillation" as the reason for one of the bans.

Distillation is a common machine-learning technique in which a smaller model is trained using outputs generated by a larger one. OpenAI itself acknowledges distillation as a legitimate method for creating training data, but its terms of use explicitly prohibit using its outputs to develop competing models or to extract output programmatically.

According to Kjellberg, his account was restored after an initial appeal, then suspended again when he ran the model to generate "seed data" for Ajax. He questioned how OpenAI detected the second attempt: "How did they even know?" he asked in the video, noting that the email offered no specific examples. OpenAI has not publicly confirmed or commented on the individual case, so the details remain based on Kjellberg's account.

Kjellberg also pushed back on the fairness of the restriction, arguing that frontier AI companies "worked hard stealing all of our data to make their AI machines," suggesting he sees little moral distinction between distillation and how large models are trained.

What "Uncensored" Actually Means

Ajax is described as "uncensored," but that label requires nuance. What Kjellberg actually did was reduce the model's refusal behavior using an open-source tool called Heretic, which performs "automatic abliteration"—a process that identifies and modifies the internal directions associated with refusals without fully retraining the system.

Kjellberg said he wanted to avoid what he called "brain damage," the risk of degrading a model's general capabilities while stripping safety guardrails. He stated that he drew the line at content that could facilitate harm to oneself or others, and that on his lawyer's advice, Ajax is "not designed to provide dangerous actionable instructions". However, these safety claims have not been independently assessed, and the actual boundaries of the model's behavior remain unclear without access to the model itself and reproducible testing.

A Bet on Small, Local AI

Ajax reflects a broader philosophical stance: Kjellberg "strongly believes" in small models trained for specific tasks rather than massive general-purpose systems. "I think it makes no sense to poke that giant beast," he said, referring to rumored trillion-parameter frontier models, arguing that such scale is unnecessary for everyday agentic tasks like browsing and email.

The local-first approach also has privacy implications. Running Ajax on personal hardware keeps data off external servers, giving users more control over their information. But as the project stands, Ajax is not yet publicly available. Its website still lists the model as "coming soon," with no published hardware requirements, downloadable weights, or independently verified benchmarks.

Kjellberg said he planned additional reinforcement learning, a second decensoring pass, quantization, and benchmarking before releasing the next version. He also put out a call for fans to donate training data as an alternative to collecting it directly from Odysseus users—though he noted that no one had submitted any at the time of the announcement.

What Comes Next

Ajax is more than a celebrity side project. It sits at the intersection of two growing trends: the push for local, private AI agents and the tension between open-source development and the restrictive terms of major AI providers. Whether Ajax ultimately delivers on its promise—and whether its "uncensored" design proves safe and useful in practice—will depend on the benchmarks, hardware support, and community adoption that have yet to materialize. For now, PewDiePie has made one thing clear: he is not content to simply use AI. He wants to build it, on his own terms.

Oh boy, I love that the guy that got caugjt paying people to engage in hate speech now has his own uncensored LLM

No way this will have any negative comments