YouTube creator Felix Kjellberg, better known as PewDiePie, has unveiled Ajax, a small language model designed to run locally on personal computers as an always-on AI assistant, following what he says were two account suspensions by OpenAI during its development.
Ajax has been developed for Odysseus, Kjellberg’s self-hosted AI workspace, and is based on Alibaba’s Qwen3.5-9B model. With around nine billion parameters, the underlying model is considerably smaller than many frontier AI systems and is intended to perform everyday tasks without relying entirely on cloud infrastructure.
The model is being developed to handle tasks including web searches, browsing, email and calendar management. Its local approach is designed to keep processing and user information on the user’s own hardware rather than routinely sending data to an external cloud service.
Kjellberg said the project also resulted in two OpenAI account suspensions. According to an email he displayed in a video announcing Ajax, one account action cited “distillation”, a machine learning technique in which outputs from a larger model can be used to help train a smaller model.
He said his account was restored following an appeal before being suspended again after he used a model to generate seed data for Ajax. OpenAI has not publicly provided a specific account of the two suspensions, meaning details about the incidents are based on Kjellberg’s statements and the communication he displayed.
OpenAI’s terms prohibit users from using outputs from its services to develop models that compete with OpenAI and restrict automatically or programmatically extracting output from its services.
Kjellberg has also modified Ajax to reduce its refusal behaviour. The project describes this process as “ablating” refusals to create a less restrictive AI experience. He has said boundaries remain around requests involving harm to users or other people. However, these safeguards have not been independently evaluated.
Ajax is still under development. Kjellberg plans further reinforcement learning, modifications to refusal behaviour, quantisation and benchmarking before a subsequent version is released.
At the time of the announcement, Ajax was listed as “Coming soon”, with downloadable model weights, final hardware requirements and independent benchmark results not yet publicly available.
The project adds to growing experimentation around smaller, specialised AI models that can operate locally. While cloud-based models continue to dominate mainstream generative AI services, local models offer an alternative architecture focused on user-controlled hardware, privacy and customisation.
Disclaimer: This article may include information derived from interviews, press releases, public statements, research, company communications and other publicly available or third-party sources. Such material may be summarised, paraphrased or contextualised for journalistic and editorial purposes. All rights in third-party content remain with their respective owners.