Pewdiepie builds smaller, local AJAX AI model after OpenAI ban controversy

YouTube creator Felix Kjellberg, better known as PewDiePie, has unveiled Ajax, a small language model designed to run locally as an always-on AI assistant. The announcement comes with an unusual backstory. Felix, who was the most-subscribed YouTuber for many years, says OpenAI suspended his account twice while he was developing Ajax, and an email shown in his launch video cites “distillation” as the reason for one of the deactivations.

Ajax is designed for Odysseus, Kjellberg’s self-hosted AI workspace, and is based on Alibaba’s Qwen3.5-9B model. The project aims to run AI locally rather than rely entirely on cloud services, with Ajax intended to handle tasks such as web searches, browsing, email, calendars, and other daily activities.

A smaller model designed for local AI

The Ajax project takes a different approach from the increasingly large frontier models powering many cloud-based AI services. Its underlying Qwen3.5-9B model has around 9 billion parameters, making it substantially smaller than the largest systems used by major AI companies.

PewDiePie argues that users do not necessarily need an enormous model for everyday agentic tasks. Instead, he is developing Ajax specifically for the Odysseus software environment, where it can interact with tools and perform practical tasks.

The Ajax project describes it as a fine-tuned model designed to be an “always-on” agent, with processing intended to remain private on the user’s hardware. However, the model’s final hardware requirements, downloadable weights, and independent benchmark results were not publicly available at the time of the announcement.

Why PewDiePie says OpenAI banned him

The more unusual part of the announcement concerns how Ajax was trained. Kjellberg says he wanted to use outputs from a larger AI model to generate training material for Ajax, a technique commonly known as knowledge distillation. In machine learning, distillation can involve using the outputs of a larger “teacher” model to train a smaller model to reproduce some of its capabilities. OpenAI itself describes distillation as a legitimate technical method for creating training data for smaller models.

The issue is how that technique is used. OpenAI’s current Terms of Use prohibit users from using its output to develop models that compete with OpenAI, as well as automatically or programmatically extracting output from its services.

According to Tom’s Hardware, Kjellberg showed an OpenAI email stating that his account had been deactivated over activity related to “distillation.” He says the account was restored after an appeal, then suspended again when he used the model to generate what he described as seed data for Ajax. OpenAIhas not publicly provided a specific explanation for Kjellberg’s two account actions, so the details of those incidents remain based on his account and the email he displayed.

Ajax is designed to be less restrictive

Kjellberg is also deliberately changing how Ajax handles refusals. The project’s website describes its refusal behavior as having been “ablated” to create what it calls a freer, less restricted AI experience.

Tom’s Hardware reports that Kjellberg used the open-source Heretic tool to perform this automatic “abliteration,” which attempts to remove some refusal behaviors from a model without extensively altering its other capabilities. He says he retained boundaries around requests involving harm to other people or oneself and, on legal advice, does not intend Ajax to provide dangerous actionable instructions.

Those safeguards have not been independently evaluated, however, so the creator’s description should not be treated as a formal safety assessment.

The model is not ready for everyone yet

Ajax remains a work in progress. Kjellberg said he planned additional reinforcement learning, followed by another round of decensoring, quantization, and benchmarking before releasing a subsequent version.

As of October 2, the project’s website was still presenting Ajax as “Coming soon,” rather than offering a public model download. There are also no published hardware requirements or confirmed license for the model weights. That makes Ajax more of an ongoing experiment than a finished competitor to commercial AI assistants for now.

Its significance may ultimately lie in the direction it represents. A recognizable internet creator is applying smaller open models, local hardware, and self-hosted software to build an AI assistant that users can potentially operate without sending their everyday tasks to a cloud-based provider.

For PewDiePie, however, the project has also become an unusual case study in the tension between that open, local approach and the restrictions imposed by the companies developing the largest AI models.