Konstantin Mishchenko announced on X on October 6 that he had joined Mistral AI’s pretraining team, which develops models through their initial large-scale training. He previously worked as a research scientist at Meta, leaving in August 2026.

He said he was excited to help develop new open models at scale and expected to publish fewer conference papers, with model quality taking priority. His research includes D-Adaptation and Prodigy, methods that reduce the manual tuning needed to train AI models.