OpenAI's layout in the field of artificial intelligence has once again made a breakthrough. Multiple tech insiders have revealed that OpenAI has successfully completed a large-scale pre-training model codenamed "Bel," with an astonishing parameter count of 10 trillion. As the next-generation base model following "Doug," Bel has achieved a qualitative leap in architecture and parameter scale, and has directly set its sights on the ultimate threshold of artificial general intelligence (AGI).
In terms of core performance and architectural capabilities, Bel outperforms the highly anticipated Astra in dimensions such as encoding, reasoning, and long-term agent tasks. According to the information, the model can run efficiently for days without human intervention, demonstrating excellent self-recovery capabilities and the ability to coordinate hundreds of parallel sub-agents simultaneously. Industry insiders point out that with the parameter count crossing the 10 trillion threshold, Bel's understanding of world models will reach an unprecedented depth, and it may even be officially unveiled by the end of this year or a few months after the release of Astra.
At the same time, OpenAI has also built a solid barrier in terms of computing power reserves and self-developed chip technology. Relying on massive computing power support and the deep integration of the underlying code of its self-developed chip "Jalapeño," OpenAI continues to lead in technical iteration speed. Through an internal recursive self-improvement mechanism, it has achieved a business and technical closed loop of using powerful models to optimize infrastructure and significantly reduce operational costs, further accelerating the evolution from cloud-based large-scale agent clusters to personal AI secretaries.
