The model is a fine-tuned version of Alibaba's Qwen3.5-9B, chosen because PewDiePie says he believes in small models trained for a specific harness rather than the rumoured trillion-parameter frontier systems, which he argues make no sense for basic tasks. To strip out refusals, he used automatic abliteration through the open-source tool Heretic, which finds the directions a model uses to refuse prompts and removes them, ideally without causing "brain damage" by cutting too much. He drew the line at harm to other people or oneself, and says on his lawyer's advice that Ajax is not designed to provide dangerous actionable instructions, though the real limits are unclear.

On the bans, PewDiePie says one OpenAI email cited "distillation", the practice of using one model's outputs or reasoning to train another, and that he was reinstated once before a second ban followed when he ran the model again to create seed data. "How did they even know?" he asked, noting the email offered no specific examples. He suggested his distillation may be little different from how the models were trained in the first place. Frontier labs such as OpenAI and Anthropic work to hide their reasoning, and OpenAI recently attributed a distillation campaign to actors linked to Moonshot AI.

Ajax is built for Odysseus, his self-hosted AI workspace launched in May, where it acts as an autonomous, always-on agent handling tasks from search and browsing to email and calendar, all locally and privately. The setup resembles OpenClaw and Nous Research's Hermes Agent. A plea for fans to donate training data drew no submissions, he says. Next up: two more days of reinforcement learning, a fresh decensoring pass, quantisation and benchmarking before the next release of Ajax and an updated Odysseus.