Introduction
PewDiePie has unveiled Ajax, a locally run AI model designed to serve as an always-on personal assistant. The project is built around a much smaller model than the frontier systems typically accessed through cloud services, with YouTube creator Felix Kjellberg describing Ajax as a fine-tuned version of Alibaba’s Qwen3.5-9B model.
The announcement also comes with a dispute involving OpenAI. Kjellberg says his OpenAI account was suspended twice while he was developing Ajax, including one ban that he says was attributed to “distillation.”
That conflict is only part of the story. Ajax is intended to power Kjellberg’s self-hosted Odysseus workspace, where it can perform tasks such as web searches, browsing, email and calendar management. But despite the launch announcement, the model is not yet publicly available, leaving several important questions unanswered about its performance, hardware requirements and eventual licensing.
PewDiePie’s Ajax Is Designed to Keep AI on the User’s Hardware
Ajax // Local AI
An AI assistant designed to run locally.
Ajax is being developed as an always-on AI assistant for local hardware, with the Odysseus environment designed to handle routine digital tasks.
Web Search
Ajax is intended to handle web searches locally through the Odysseus environment, reducing the need to rely on a separate cloud-based assistant for routine search tasks.
The project still requires further training and benchmarking before public availability. The project site currently describes it as “coming soon.” Hardware requirements and a confirmed model-weight license have not been published.
Ajax Uses a 9B-Parameter Qwen3.5 Base
The technical foundation of Ajax is considerably smaller than the frontier models powering many major commercial AI services.
Kjellberg describes it as a fine-tuned version of Alibaba’s Qwen3.5-9B, a model with roughly 9 billion parameters.
That size makes Ajax interesting for local AI experimentation because the project is targeting a model that can potentially be deployed on consumer hardware rather than depending exclusively on a large cloud infrastructure.
But the announcement does not establish what hardware will ultimately be required.
There is currently no published hardware specification in the supplied material, so it would be premature to assume that Ajax will run comfortably on any particular laptop, desktop or GPU.
Its actual local performance will need to be established through the benchmarking Kjellberg says is still ahead.
Why OpenAI Became Part of the Ajax Story
The most controversial part of the announcement concerns Kjellberg’s use of OpenAI systems during development.
Kjellberg says OpenAI suspended his account twice while he was working on Ajax. He showed an email in his launch video that reportedly cited “distillation” as the reason for one of the bans.
The issue is particularly relevant because distillation itself is not inherently an unusual machine-learning technique.
In broad terms, model distillation involves using the outputs of a larger model to help train a smaller model. The technique can be used to transfer some capabilities from a larger system into a more compact one.
According to the supplied reporting, OpenAI permits the technique in some contexts but its terms prohibit using OpenAI outputs to build competing models or extracting those outputs programmatically.
Kjellberg says he wanted to “distil just a little bit” to improve Ajax and questioned how OpenAI identified his second attempt.
The dispute therefore centers less on whether distillation exists as a legitimate AI technique and more on how a particular use of it interacts with OpenAI’s rules.
Ajax Also Changes How the Model Handles Refusals
Kjellberg says he modified Ajax’s refusal behavior with Heretic, an open-source tool designed to remove refusal responses while attempting to preserve other model capabilities.
That is another area where the project needs careful qualification.
Kjellberg says Ajax will retain restrictions around self-harm and harming other people. He also says legal advice influenced his decision not to provide dangerous instructions.
Those safeguards, however, have not been independently assessed, according to the supplied reporting.
That means Ajax’s eventual safety behavior should not be treated as independently verified simply because Kjellberg describes specific restrictions he intends to retain.
Ajax Is More Ambitious Than a Local Chatbot
The project’s intended role extends beyond simply answering questions.
Ajax is supposed to sit inside Odysseus as an always-on assistant capable of interacting with everyday digital workflows.
That distinction matters.
A local language model that only generates text is one thing. An assistant that can search the web, browse websites, interact with email and manage calendars requires an additional layer of tools and permissions around the model.
In other words, the model itself is only one component of the system Kjellberg is building.
The eventual usefulness of Ajax will depend not only on its language-model performance but also on how reliably that surrounding workspace can execute tasks.
Ajax Is Still Not Available to the Public
Despite the announcement, readers cannot download and test Ajax yet.
Kjellberg says the project still needs:
- Further training
- Decensoring work
- Benchmarking
The project site also lists Ajax as coming soon.
Several practical details remain unavailable, including the model’s final hardware requirements and a confirmed license for its weights.
That makes it too early to compare Ajax directly with established local AI models on performance, efficiency or compatibility.
For now, the announcement establishes the project’s direction rather than delivering a finished public release.
What PewDiePie’s Project Could Tell Us About Local AI
The more interesting development may be the type of AI assistant Ajax is attempting to become.
The project combines a relatively compact language model with a self-hosted workspace and access to everyday digital tools. If the finished system performs as intended, that could make local AI more useful as a persistent personal computing layer rather than simply an offline chatbot.
But that remains an objective, not a demonstrated result.
Ajax still needs additional training and benchmarking, and the public release does not yet provide enough information to judge its real-world performance.
The next meaningful milestone is therefore not another claim about what the model could theoretically do. It is evidence from the finished model itself: benchmarks, hardware requirements, licensing details and eventually public access.
What Comes Next for Ajax
The immediate question is when Ajax will move from an announced project to a model that others can actually evaluate.
Kjellberg has indicated that more training, decensoring and benchmarking are still ahead. The absence of published hardware requirements and a confirmed model-weight license also leaves important practical questions open.
For readers interested in local AI, those details will matter more than the project’s celebrity association.
Ajax has been unveiled, but it has not yet been proven. The next stage will determine whether PewDiePie’s experiment becomes a genuinely useful local AI assistant or remains an ambitious work in progress.
Practical Takeaway
PewDiePie’s Ajax is best understood as an unfinished local AI project, not a publicly available replacement for commercial AI assistants.
The important facts are:
- Model: Fine-tuned Qwen3.5-9B
- Scale: Roughly 9 billion parameters
- Purpose: Always-on local AI assistant
- Workspace: Odysseus
- Planned capabilities: Search, browsing, email, calendars and other daily tasks
- OpenAI dispute: Kjellberg says his account was suspended twice during development
- Current status: Not publicly available
- Still needed: Further training, decensoring and benchmarking
- Unknown: Final hardware requirements and confirmed model-weight license
The next useful development will be the point at which Ajax can be independently tested rather than simply described by its creator.
What You Need to Know About Ajax
Tap a question to uncover the answer.
Ajax is a locally run AI model unveiled by PewDiePie. It is designed to work as an always-on assistant through his self-hosted Odysseus workspace rather than relying entirely on cloud-based AI services.
PewDiePie describes Ajax as a fine-tuned version of Alibaba’s Qwen3.5-9B model, which contains roughly 9 billion parameters.
The planned assistant is designed to handle everyday digital tasks such as web searches, browsing, email and calendar management. These capabilities are part of the project’s intended functionality and are still being developed.
Kjellberg says OpenAI suspended his account twice while he was developing Ajax. An email shown in his launch video cited “distillation” as the reason for one of the suspensions.
Model distillation is a technique in which outputs from a larger model can be used to help train or improve a smaller model. The controversy here concerns how such use interacts with OpenAI’s terms rather than whether distillation itself is a recognized machine-learning technique.
Kjellberg says he used the open-source Heretic tool to alter Ajax’s refusal behavior. He also says safeguards around self-harm and harming others remain in place. Those safeguards have not been independently assessed.
No. Ajax is not publicly available yet. Further training, decensoring and benchmarking are still planned, while its final hardware requirements and model-weight license have not been confirmed.
A local model can perform its AI processing on the user’s own hardware rather than depending entirely on remote cloud infrastructure. Ajax is particularly interesting because it aims to combine that local approach with tools for everyday digital tasks.
Loop Teck
✓Technology, AI, smartphones & the future of innovation.
