The Ajax AI model is a fine-tuned version of the Qwen3.5-9B architecture, designed to operate locally on home computers. It aims to manage daily tasks such as web browsing, email, and calendar management while ensuring user privacy. The developer removed the model's refusal mechanisms to provide a less restricted experience, though it remains programmed to avoid generating dangerous instructions.
The creator utilized the open-source tool Heretic to perform automatic abliteration, which involves identifying and removing prompts that trigger refusals. This process is intended to be precise to avoid degrading the model's performance. The project follows the launch of a workspace initiative introduced in May and is intended as a compact alternative to massive models with trillions of parameters.
The development process involved challenges with platform access, as the creator reported being banned by OpenAI twice during the model's creation. He noted that the platform provided no specific reasons for these restrictions. The developer expressed skepticism regarding the data collection practices of large AI companies, comparing his own distillation methods to industry-standard training techniques.
Future plans for the project include two additional days of reinforcement learning, followed by quantization and benchmarking. The creator is also working on updates for the Odysseus workspace and has requested that users contribute training data to improve the model. He emphasizes the importance of maintaining access to uncensored, open-source intelligence that runs directly on consumer hardware.
