36Kr Reports OpenAI Model 'Bel' Finished Pre-Training, Claims 10T Parameters

In one line: Chinese outlet 36Kr reports that a next OpenAI model codenamed 'Bel' has finished pre-training at a claimed 10 trillion parameters — but there is no official confirmation.

Key points

• 36Kr says a new OpenAI model, referred to as 'Bel,' has completed the pre-training stage.
• The report cites a scale of 10 trillion parameters, which, if accurate, would rank among the largest publicly discussed models.
• No OpenAI announcement or technical documentation has corroborated the claim, so both the figure and the model name rest on a single outlet's reporting.

Why it matters

Parameter count is often used as a proxy for training cost, infrastructure needs, and expected capability. Yet the recent industry trend leans toward data quality and inference efficiency over raw parameter growth, making actual performance and public availability more decisive than the headline number. Until OpenAI confirms it, the claim is best treated as unverified.

Read more

• Breaking News: ChatGPT's New Model "Bel" Completes Pre-Training — 36Kr (via Google News)