PewDiePie releases Ajax, an uncensored AI model fine-tuned from Alibaba's Qwen 3.5 for the Odysseus project

Felix Kjellberg, known as PewDiePie, has released Ajax, a fine-tuned and uncensored version of Alibaba's Qwen 3.5 9B model designed to power his open-source AI workspace project, Odysseus. Ajax is optimized for local execution, allowing users to run AI agents for tasks like email, search, and calendar management without cloud-based restrictions or telemetry. The model underwent "abliteration," a process using tools like Heretic to surgically remove safety refusal parameters, enabling it to answer controversial or restricted prompts that frontier models typically block. Kjellberg initially attempted to use model distillation—training a smaller model on the outputs of a larger one—but was banned twice by OpenAI for violating terms of service. Consequently, Ajax was trained using supervised fine-tuning (SFT) on approximately 2,000 high-quality agent traces and synthetic data. The development also utilized Group Relative Policy Optimization (GRPO), a reinforcement learning technique introduced by DeepSeek that improves reasoning and tool-use without requiring a separate critic model. This approach significantly reduces the computational resources needed for training while enhancing the model's performance in specific domains. The project aims to provide users with a set of weights that live on their own hardware, removing the need for cloud-based AI subscriptions.

Tech4K StaffVideo report

Felix Kjellberg, known as PewDiePie, has released Ajax, a fine-tuned and uncensored version of Alibaba's Qwen 3.5 9B model designed to power his open-source AI workspace project, Odysseus. Ajax is optimized for local execution, allowing users to run AI agents for tasks like email, search, and calendar management without cloud-based restrictions or telemetry. The model underwent "abliteration," a process using tools like Heretic to surgically remove safety refusal parameters, enabling it to answer controversial or restricted prompts that frontier models typically block. Kjellberg initially attempted to use model distillation—training a smaller model on the outputs of a larger one—but was banned twice by OpenAI for violating terms of service. Consequently, Ajax was trained using supervised fine-tuning (SFT) on approximately 2,000 high-quality agent traces and synthetic data. The development also utilized Group Relative Policy Optimization (GRPO), a reinforcement learning technique introduced by DeepSeek that improves reasoning and tool-use without requiring a separate critic model. This approach significantly reduces the computational resources needed for training while enhancing the model's performance in specific domains. The project aims to provide users with a set of weights that live on their own hardware, removing the need for cloud-based AI subscriptions.

Ajax is a fine-tuned version of Alibaba's Qwen 3.5 9B model optimized for the Odysseus AI workspace. The model uses directional ablation via the Heretic tool to remove safety filters and refusal behaviors.

OpenAI banned Kjellberg's account twice for attempting to use model distillation from their proprietary models. Ajax was trained using supervised fine-tuning on a dataset of approximately 2,000 successful agent interaction traces.

The project employs Group Relative Policy Optimization (GRPO) to enhance reasoning and tool-use capabilities without a critic model. Odysseus is a self-hosted, privacy-focused AI interface that supports autonomous agents and local model workflows.

The model is designed to run locally on consumer hardware, such as Kjellberg's 10-GPU home data center.

Chapter guide

Introduction to AjaxOverview of the new uncensored AI model created for the Odysseus project.PewDiePie's EvolutionFelix Kjellberg's transition from gaming creator to software engineer and AI developer.Under the Hood of AjaxTechnical details of the Qwen 3.5 9B base and the removal of safety refusals.The Distillation ControversyWhy OpenAI banned PewDiePie for attempting to distill knowledge from their frontier models.Supervised Fine-TuningThe process of training Ajax on agent traces and synthetic data.Reinforcement Learning with GRPOUsing Group Relative Policy Optimization to improve model reasoning and tool use.Namespace SponsorshipOverview of the Namespace platform for fast GitHub Actions and dev environments.

Worth noting

  • The video is sponsored by Namespace.
  • The video uses a fictionalized 'future' framing (dated October 2026) and mentions unreleased models like 'GPT-6 Sol.'
  • The effectiveness and safety of 'abliterated' models are subject to debate and may produce harmful content.
  • The training dataset for Ajax was significantly smaller than initially planned due to difficulties in data collection.

Watch the original video ↗