PewDiePie Says OpenAI Banned Him Twice While Building His “Uncensored” AI

PewDiePie has released an AI model designed to run locally on consumer hardware, giving his self-hosted AI workspace a model built specifically for autonomous tasks rather than ordinary chatbot conversations.

“Qwen3.5-9B” is a 9-billion-parameter model, according to the official Qwen model documentation, which also lists support for local deployment through frameworks including Transformers, vLLM and SGLang.

“Ajax is a fine-tuned AI model designed specifically to work with Odysseus,” Kjellberg said in his launch video, describing capabilities including web searches, private searches, email and calendar assistance. The video also presents the model as a small local system intended to run on the user’s own computer.

What “Uncensored” Means In Ajax

Kjellberg describes Ajax as “uncensored,” but the term has a narrower meaning than it might suggest.

“Choose a subset of modules to modify, then choose how to modify each of them, with the goal of suppressing refusals as much as possible, while retaining as much model intelligence as possible,” Heretic creator p-e-w explains in the project’s documentation.

That is Kjellberg’s description, not the result of an independent safety evaluation. At launch, Ajax did not have a public model card or published benchmark results, and its weights were not publicly available.

The project’s download link instead redirected to a page collecting training data. Kjellberg has also reported an approximately 90% task-completion rate in early testing inside Odysseus, although he has described the project as a work in progress.

Kjellberg Says OpenAI Suspended His Account

In his unveiling video, Kjellberg showed an email saying his account had been deactivated over activity related to “Distillation.” He said he appealed and regained access, but was later suspended again after using an OpenAI model to generate what he described as seed training data, including chain-of-thought outputs from OpenAI’s Sol model. Kjellberg questioned how OpenAI detected the activity.

“This activity is consistent with adversarial distillation: the systematic and unauthorized use of one model’s outputs or reasoning to help train, reproduce, or improve another model,” OpenAI said in its Sept. 30 account of the campaign it disrupted.

There is relevant context, but it is a separate incident. On Sept. 30, OpenAI said it had disrupted a campaign it attributed to associates of Moonshot AI that attempted to extract protected reasoning outputs at scale. OpenAI described that activity as adversarial distillation. The parties and circumstances are different, but the timing shows that model distillation was an active enforcement issue for OpenAI when Kjellberg’s video appeared.

Why A 9-Billion-Parameter Local Agent Matters

“Qwen3.5 excels in tool calling capabilities,” the Qwen team says, recommending Qwen-Agent for building agent applications. The official model documentation also provides instructions for running Qwen3.5-9B locally through several inference frameworks.

That points toward a different model of personal AI: instead of sending every request and piece of context to a cloud provider, an agent can keep its model and potentially much of its working data on the user’s own machine.

The idea itself is not new. Open-weight models such as Qwen have already made local AI increasingly accessible, and developers have been building local agents around them. What makes Ajax notable is the audience behind it. Kjellberg is putting the self-hosted agent concept in front of a mainstream audience that may never have encountered local AI through developer communities.

But Ajax is not yet a fully documented public release. Without publicly available weights, a model card and independent benchmark results, its claimed performance and safety characteristics cannot yet be evaluated in the same way as established open-weight models.