Elon Musk promised to release a new version of Grok every two weeks, but the three core team members building the model all resigned within a week.

According to Beating Monitoring, over the past week, three technical core figures directly in charge of building the Grok model have, one after another, announced their resignations at xAI. Unlike the earlier exits involving the co-founders and senior executives, these three are the actual heads of the three core product lines—post-search training, pre-training, and speech. Their collective departure will directly affect the iteration capability of Grok’s next-generation models.

Tianyi Zhang confirmed his departure today on X. He is the head of the search and factual post-training team; previously, he served as a machine learning engineering manager at Apple for five years. Under his leadership, the factual error rate of Grok’s real-time mode was reduced by 70% to 80%, helping Grok 4.20 top the Search Arena blind test leaderboard immediately upon launch. This search capability has also been integrated into Tesla’s in-car system.

Juntang Zhuang announced on May 9 that he had already left earlier this year. He has been deeply involved in pre-training since Grok 2, later growing into the head of the pre-training team, where he led the training infrastructure and core recipes for the entire Grok 4 series of models. Before joining xAI, he worked at OpenAI for two years, as a core contributor to GPT-4o and the inventor of the GPT-4 Turbo 128k long-context algorithm.

David Haxton confirmed his resignation on May 10. Previously, he built Grok’s speech agent, real-time thinking speech, and voice cloning functions from scratch. In Sierra AI’s τ-Voice speech agent benchmark test, Grok Voice, the team he led, is currently ranked first—ahead of OpenAI and Google.

The timing of these resignations is especially striking. As of now, the publicly available Grok 4.3 is still only an interim training version with 0.5T parameters. On April 18, Musk said the true 1T flagship would be “about 5 days away from initial training completion,” and the next day he further clarified that Grok 4.4 = 1T, expected to be released in early May; Grok 4.5 = 1.5T, with delivery targeted for the end of May; he also claimed that the model “factory” had been running smoothly, and that a new base model would be taken offline every two weeks thereafter. Now it is mid-May, and Grok 4.4 has yet to show up, while the key personnel who actually built these models are accelerating their departures.

xAI is undergoing a systematic talent collapse—from co-founders to the core layer building the models.