Skip to main content

OpenAI's 10-Trillion-Parameter Model 'Bel' Finishes Pre-training

OpenAI has once again pushed the boundaries of artificial intelligence. According to multiple tech insiders, the company has successfully completed pre-training for a massive new model codenamed "Bel," boasting a staggering 10 trillion parameters. As the successor to the "Doug" model, Bel represents a qualitative leap in both architecture and scale, and it's squarely aimed at the ultimate prize: artificial general intelligence (AGI).

Image

What makes Bel stand out? For starters, it outperforms the highly anticipated Astra in key areas like coding, reasoning, and long-horizon agent tasks. The model can run efficiently for days without human intervention, showcasing impressive self-recovery capabilities. It can also coordinate hundreds of parallel sub-agents simultaneously—a feat that hints at a new level of autonomous operation.

Industry observers note that crossing the 10-trillion-parameter threshold means Bel's understanding of world models will reach unprecedented depth. Some speculate it could be officially unveiled by the end of this year or a few months after Astra's release.

But Bel isn't just about raw size. OpenAI has also fortified its position with massive computing power and self-developed chip technology. The deep integration of its custom "Jalapeño" chip's underlying code gives the company a speed advantage in technical iteration. Through an internal recursive self-improvement mechanism, OpenAI has created a closed loop: powerful models optimize infrastructure, which in turn reduces operational costs. This accelerates the evolution from cloud-based agent clusters to personal AI assistants.

While OpenAI hasn't officially confirmed these details, the buzz is undeniable. If Bel lives up to expectations, we could be looking at a significant milestone on the road to AGI.

Key Points

  • Bel's Scale: 10 trillion parameters, making it one of the largest models ever.
  • Performance: Outperforms Astra in coding, reasoning, and long-term tasks.
  • Autonomy: Runs for days without human help, coordinates hundreds of sub-agents.
  • Infrastructure: Leverages custom Jalapeño chips and recursive self-improvement.
  • Timeline: Possible release by end of 2026 or shortly after Astra.