PewDiePie says OpenAI banned his account twice while he was building Ajax, his new “uncensored” AI model meant to run on people’s own computers, and that one of the bans was explicitly for “distillation.” The YouTuber, whose real name is Felix Kjellberg, unveiled the Ajax AI model in a new video posted this week, as Tom’s Hardware reported on Oct. 2, 2026.
There is a catch for anyone hoping to try it tonight: Ajax is not out yet. Its official page currently reads “Ajax — Coming soon,” and Kjellberg has not published hardware requirements, a model card or a license for the weights.
What the Ajax AI model actually is
According to its official page at data.pewdiepie.com, “Ajax is a fine-tuned Qwen 3.5 9B model, trained for Odysseus to be your always on agent.” In plain terms, Kjellberg took an existing open model from Alibaba, Qwen3.5-9B, which has 9 billion parameters, and trained it further for one specific job.
That job is powering the Odysseus workspace, the self-hosted AI app he released earlier this year. The Ajax page says the model “handles your daily tasks from search to browse web to email to your calendar – all your daily tasks completely privately.”
Odysseus itself is very real and very popular. Its GitHub repository, now under the odysseus-dev organization, describes it as “A self-hosted AI workspace for chat, agents, research, documents, email, notes, calendar, and local model workflows.” It is licensed under AGPL-3.0-or-later and showed about 88,600 stars when we checked on Oct. 2.
The OpenAI ban, in his words
The part of the video that took off online is the account trouble. Tom’s Hardware, which viewed the video, reports that the OpenAI email Kjellberg shows on screen cites “distillation,” the practice of using one model’s outputs or reasoning to train another model.
Kjellberg did not deny what he was trying to do. He said he wanted to “distill just a little bit” to improve his model. He got the account reinstated once. Then, after he ran OpenAI’s model again “to create my seed data,” he was banned a second time.
“How did they even know?”
That was his reaction to the second ban, according to Tom’s Hardware, which notes the email offers no specific examples. He also took a swipe at the company: “They worked hard stealing all of our data to make their AI machines.” The line echoes a broader argument about how AI companies built their models in the first place, a debate NarwhalTV covered when an AI executive’s “theft of labor” remark went viral.
Has OpenAI responded?
Not publicly. We found no statement from OpenAI about Kjellberg’s account, and the only public account of both bans is his own video. OpenAI’s consumer terms do bar using its outputs to develop competing models, according to coverage of the case, and the company has recently been aggressive about model distillation. Tom’s Hardware points to a recent campaign OpenAI attributed to actors linked to Chinese AI company Moonshot AI, which tried to extract the hidden reasoning of its models. Nothing has been reported linking that episode to Kjellberg.
How the “uncensored” part works
The Ajax page says the model’s “refusal has been ablated for a freer, less restricted AI experience. Please use responsibly.” Kjellberg used an open-source tool called Heretic to do this automatically. The method, known as abliteration, finds the internal patterns that make a model refuse certain prompts and strips them out, ideally without causing what he called “brain damage” by removing too much.
He said he drew the line at harming other people or oneself. On his lawyer’s advice, he said Ajax is “not designed to provide dangerous actionable instructions,” using nuclear weapons as his example. Tom’s Hardware notes it is unclear where the real limits fall, and no public testing exists yet to check.
Running Qwen3.5-9B at home
The pitch is a small model that does a few things well on a home machine instead of calling a giant cloud system. Kjellberg said he “strongly believe[s]” in small models trained “for a specific harness,” and argued against using “one trillion parameter” models for basic chores. “I think it makes no sense to poke that giant beast,” he said.
A 9-billion-parameter model is small by frontier standards, and in a compressed, or quantized, form it can fit on many gaming graphics cards. But the exact requirements for Ajax will depend on which formats he releases. For now, the claim that it is “built to run on home PCs” is a design goal rather than something buyers can verify.
The local, private approach is also part of why self-hosted AI tools have gained fans. Cloud assistants that act on your behalf have produced their share of personal AI agent horror stories, and keeping data on your own desk is one answer to that worry.
Release plans for the Ajax AI model
There is no launch date. The Ajax page now says: “I’ve decided to release Ajax when it’s ready instead of putting a timer. I’m sorry, I’ll do my best to get it done as soon as possible!”
In the video, he said he planned two more days of reinforcement learning, followed by rerunning the decensoring step, quantizing and benchmarking. “I can’t wait to release the next version of Ajax and Odysseus,” he said, adding that Odysseus has “gotten a huge update as well.”
He also asked fans to donate training data rather than collect it from Odysseus users, though he said no one had submitted any. As for the name, Ajax was a rival of Odysseus in Greek myth. Kjellberg joked that it is named after the cleaning product.
Until the weights are public, the key claims, including how well Ajax performs and how far its guardrails really go, remain his word alone.
