Three Core Grok Model Architects Depart xAI Within One Week as Musk Promises Biweekly Releases

According to monitoring by Beating, three technical leads directly responsible for Grok model development departed xAI within the past week. Tianyi Zhang, who led search and factuality post-training, confirmed his departure today; Juntang Zhuang, pre-training team lead, left earlier this month; and David Haxton, who built Grok’s voice agent capabilities, departed May 10.

These exits target core product lines—search post-training, pre-training, and voice—directly responsible for Grok’s next-generation iteration capacity. Zhang’s team reduced Grok’s factual error rate by 70-80%, enabling Grok 4.2 to top the Search Arena leaderboard. The timing compounds pressure: Musk claimed in April that Grok 4.4 (1T parameters) would launch in early May with biweekly releases thereafter, yet remains unannounced. Grok 4.3, currently public, is only a 0.5T training intermediate version.