1. Home
2. News
3. OpenAI

PewDiePie built Ajax, a small home AI model, and says OpenAI banned him twice

The download link in his video goes to a page that says the model isn't out yet.

By Think Facility · October 4, 2026 · OpenAI Local AI Save

PewDiePie announced Ajax, a small AI model he fine-tuned to run on home computers, in a YouTube video posted late on September 30, 2026, US time. In the same video he says OpenAI banned his account twice while he was building it.

The video's description says "Download Ajax here" and links to data.pewdiepie.com. That page doesn't offer a download. It says he'll release the model once it's finished.

Model

a fine-tuned Qwen 3.5 9B model

Built for

Odysseus, his self-hosted AI workspace

Refusals

Ajax's refusal has been ablated

Release

Ajax when it's ready

What Ajax is for

Ajax is built to drive Odysseus, the open-source workspace PewDiePie's GitHub account maintains. Odysseus runs chat, agents, email and research against models on your own machine. Its repository has about 89,400 stars.

The base is Alibaba's Qwen3.5-9B. He picked a 9-billion-parameter model on purpose, and explained why on GitHub on June 5, 2026, when he opened a discussion asking users for help with a dedicated Odysseus model.

In the video he says the first version gets its tasks right about nine times out of ten. He also does some napkin math: running one of the rumored trillion-parameter frontier models would take roughly 27 of his own computers, which he thinks is absurd for checking an email.

The two bans

To make Ajax smarter, he says, he wanted to distill a little from OpenAI's GPT-6.1 Sol (distillation means training a small model on a larger one's outputs). OpenAI banned him. He appealed, and the account came back. Then he ran Sol again to make seed data for training, and OpenAI banned him a second time. He asks, loudly, how they even knew.

He doesn't pretend to be sorry. He jokes about a method for reading Sol's hidden reasoning, then says he'd never use it because it breaks the terms of service, and points out that the big labs built their models on everyone else's data. We didn't find an OpenAI post about his account.

The refusals

The Ajax page says the model's refusal behavior was removed. In the video he says he drew a line at harm to other people and to yourself, and that his lawyer told him to say Ajax isn't designed to give dangerous, actionable instructions (nuclear weapons are his example).

He's still training. The video says he's been running reinforcement learning for weeks, and that the plan is two more days of it, then removing refusals again, quantizing the model and benchmarking it. He also asked viewers to donate Odysseus usage data through the same site. By his account, nobody has yet.