OpenAI has reportedly completed a pretraining run named 'Bel' that used more than 10 trillion parameters. The information comes from a report that has not been independently verified, and the company has not publicly confirmed it.
What the report says
The report says the training run is finished. That's about all it says. No details on the model's purpose, the data used, or the timeline of the run have been shared. The name 'Bel' appears to be an internal code name, but nothing about it is explained.
At its core, this is a single claim: a pretraining run happened, and it involved an extraordinarily large number of parameters.
What pretraining means
Pretraining is the first major step in building a large language model. The model is fed a massive amount of text, learning patterns, grammar, and relationships between words. The number of parameters is a common way to measure a model's size and capacity. More parameters generally mean the model can store more information and handle more complex tasks, but they also demand far more computational power during training.
Ten trillion parameters is a scale that pushes far beyond what most publicly known models use. But that comparison is based on general knowledge, not on any specific figure in the report.
What this scale would take
If the report is accurate, training a model of this size would require a vast array of specialized processors, a carefully designed data pipeline, and a long, continuous operation. Such an effort would come with a significant cost, though no dollar figure is mentioned. It also means the training data would have to be enormous — without a huge dataset, a model with 10 trillion parameters would have little to learn from.
The report doesn't say whether this run is an experiment in scale or a step toward a product. It doesn't say whether the resulting model is meant for internal use or public release.
What's missing
OpenAI has not commented on the report. The company's usual practice of staying quiet about internal research makes it hard to know if the claim is real, exaggerated, or simply a rumor. No independent party has confirmed the run. The report itself doesn't name a source beyond the claim.
The name 'Bel' is the only concrete detail beyond the parameter count. That doesn't reveal much — no explanation is given for why it's called that, or what the model is expected to do.
There are no specifics on the training data, the duration of the run, or the hardware used. All of those remain blank.
The open question
The only fact that exists is the report itself. Whether OpenAI actually completed this run, and what it plans to do with the model, is still unknown. The next step would be a confirmation or denial from the company — but so far, there's no indication that one is coming.
Until then, this is a one-line claim with no supporting detail.
