Felix Kjellberg, better known as PewDiePie, has unveiled the Ajax AI model, a fine-tuned 9-billion-parameter assistant designed to run locally through his Odysseus workspace. He also says OpenAI suspended his account twice while he was developing it.
In his launch video, Kjellberg says Ajax can search and browse the web, handle email, manage a calendar and work with to-do notes. Tom’s Hardware reports that the model is based on Alibaba’s Qwen3.5-9B and fine-tuned for Odysseus, a self-hosted AI workspace.
What the Ajax AI model is built to do
The Ajax AI model is meant to use Odysseus’s tools rather than operate as a standalone chatbot. That puts the focus on carrying out small jobs inside a user’s own workspace, such as finding information or organising a task, rather than trying to match a cloud model across every kind of prompt.
Kjellberg said the early version completed its intended tasks roughly nine times out of ten. He also said he had continued training it after that version, including a reinforcement-learning stage, and described it as a work in progress.
The software and the model are separate products. The Odysseus repository describes a self-hosted workspace and lists its own licence as AGPL-3.0-or-later; that does not establish the licence or redistribution terms for Ajax’s weights.
As of 2 October, the model files are not available from the link in the video description. Although it is labelled “Download Ajax here”, it opens a page asking users to contribute training data to Odysseus, not a model download. Free Press Journal also reported on 2 October that Ajax had no public download, final model card or confirmed weights licence at that time.
Kjellberg has not published hardware requirements or supported runtimes for Ajax. Tbreak’s guide to memory for local AI on Mac explains the variables buyers need to consider when deciding whether a local model will fit their machine.
What PewDiePie says about OpenAI
Kjellberg says OpenAI banned his account twice while he was trying to create training data for Ajax from model outputs. He showed an email in the video that says his account had been deactivated over activity related to “Distillation”; he says it was restored after an appeal, then suspended again after he ran OpenAI’s model to create seed data.
He also discusses a research method for extracting more of a model’s reasoning, but says he would not use it because it would breach the service’s terms. OpenAI’s Terms of Use prohibit using its output to develop models that compete with OpenAI. That general rule does not confirm what prompted either account decision, and the video and reports reviewed do not include an OpenAI explanation of the bans.
Free Press Journal likewise attributes the account history to Kjellberg, rather than to a statement from OpenAI. The first email he showed supports his account of one suspension; the reasons for the second remain his account of events.
Fewer refusals, with safety limits still untested
Ajax is also designed to refuse fewer prompts. Tom’s Hardware says the project page describes its refusal behaviour as ablated for a “freer, less restricted” experience. Kjellberg says he used the open-source Heretic tool and chose to retain limits around requests to harm other people or oneself.
He says Ajax is not designed to provide dangerous, actionable instructions. No public evaluation shows how consistently those limits work, however, so the creator’s stated boundary should not be mistaken for an independently tested safety result. The available material also does not explain the model’s permission controls when it is connected to email, calendar or web tools.
Whether Ajax is useful outside its creator’s setup will depend on downloadable weights, supported runtimes and hardware guidance, as well as the quality of its tool use. Tbreak’s Mac Studio M5 Max review looks at the hardware side of running AI locally, but Ajax’s own requirements remain unknown.
Does a 9B parameter count mean Ajax will run on an ordinary laptop?
No. Parameter count does not tell you the memory or storage needed for a particular build. Usage depends on weight precision, quantisation, context length and runtime, so users should wait for downloadable files and official hardware requirements before deciding.
Does PewDiePie’s nine-in-ten task result prove Ajax is reliable?
No. It is a result Kjellberg reported for his own early tests, without a disclosed task set, sample size or independent comparison. It cannot be compared with published benchmark scores.
Does Odysseus’s AGPL licence also cover Ajax’s model weights?
Not automatically. The repository identifies the Odysseus workspace software as AGPL-3.0-or-later, while fine-tuned model weights can have separate terms. Ajax’s licence and redistribution conditions need to be published separately.
Does running Ajax locally mean every connected task stays on the computer?
Not necessarily. Model inference may run locally, while browsing and connected email or calendar tools still interact with external services. Users need to check connector permissions and data flows before linking personal accounts; the current material does not document those controls.
