PewDiePie Says OpenAI Suspended Him Twice While He Built His Ajax AI Model
Key Takeaways
- •PewDiePie, whose real name is Felix Kjellberg, released Ajax, a 9-billion-parameter AI model fine-tuned from Alibaba's Qwen3.5-9B base model.
- •Kjellberg claims OpenAI suspended his account twice during Ajax's development, with the alleged reason being distillation of OpenAI outputs, including one referred to as 'Sol,' into training data.
- •Ajax is designed to run locally on consumer hardware as an autonomous agent within Odysseus, the self-hosted workspace Kjellberg launched in May 2026, covering tasks like web searches, email, calendars, and to-do lists.
- •The model's tendency to refuse requests was reduced through 'abliteration' using a tool called Heretic, though certain safety bounds against harmful requests reportedly remain in place.
- •As of early October 2026, Ajax lacked a public model card, published benchmarks, a confirmed weights license, and a download link.

Felix Kjellberg, the YouTube personality known to millions as PewDiePie, has released his own artificial intelligence model, which he calls Ajax. According to Kjellberg, building it got him kicked off OpenAI not once, but twice.
The model arrived around October 1–2, 2026. It is a small, uncensored assistant designed to live on a user's own computer rather than on a company's server farm.
What Ajax Is
Ajax is a 9-billion-parameter model, and Kjellberg did not build it from scratch. It is a fine-tuned version of Alibaba's Qwen3.5-9B base model. Fine-tuning means taking an existing base model and training it further to shape its behavior — a route most small, independent AI projects take rather than paying to train a model from the ground up.
The goal is practical. Ajax is meant to act as an autonomous agent and assistant inside Odysseus, a self-hosted workspace Kjellberg launched in May 2026. Its task list is deliberately unglamorous: web searches, email management, calendar duties and to-do lists. All of it is designed to run on consumer hardware, meaning searches and files can stay on the user's own machine instead of passing through a provider's servers.
The Two Suspensions
The launch came with its share of drama. Kjellberg claims OpenAI suspended his account on two separate occasions while he was developing Ajax. The stated reason, according to his account, was distillation — the practice of feeding a powerful model's answers into a smaller model as training material. The technique is common across machine learning, and how provider terms treat it has become a recurring point of contention as more builders turn larger models' outputs into training data for smaller ones.
In this case, the allegation centers on outputs from OpenAI models, including one referred to as “Sol,” being used to generate training data for Ajax.
The Uncensored Part
Ajax also leans on a tool called Heretic for a process known as “abliteration.” Abliteration is a technique for removing a model's tendency to refuse requests, stripping out much of that reflex so the model simply answers.
According to the research findings, Ajax still keeps certain safety bounds in place against harmful requests. Uncensored here, then, appears to mean less preachy rather than entirely unrestricted.
What Is Still Missing
As of early October 2026, several standard pieces of a model release were absent. Ajax had no public model card — the customary document spelling out what a model does, how it was built and where it falls short — no published benchmarks, no confirmed weights license, the terms under which others may download and reuse the model itself, and no immediate download link.
Background: A YouTuber Goes Self-Hosted
Kjellberg's AI project did not start with Ajax. Odysseus came first in May 2026, providing a self-hosted environment for an assistant to operate in. Ajax is effectively the brain he built for that body.
What This Means
Kjellberg's account, if accurate, shows OpenAI's terms being enforced against a high-profile individual developer, not just a rival company. The research findings suggest interest in privately run, open-source AI systems continues to surge, and a creator with Kjellberg's reach shipping a home-hardware assistant gives that idea a very public showcase.
An abliterated model with limited documentation invites scrutiny over how its safety bounds actually hold up. A model reportedly trained partly on another lab's outputs could also face questions about what its eventual license can legitimately permit.
A published model card and benchmarks would show whether Ajax performs as advertised. A confirmed license and download link would reveal whether it becomes a genuine open tool or remains a showcase.
Source: CryptoBriefing