Leaving Meta and PyTorch (soumith.ch)

Soumith Chintala announced he is leaving Meta and stepping away from day-to-day leadership of PyTorch after 11 years at Meta and nearly eight years growing PyTorch into the dominant ML framework. He highlights PyTorch’s leap to exascale training, its role powering foundation models and production systems across major AI companies, and its 90%+ adoption in the AI community—from top universities to under-resourced classrooms. Soumith framed this as a deliberate exit: planned during parental leave, timed because the project and leadership bench (Edward, Suo, Alban, Greg, John, Joe, Jana and others) are ready, and because he wants to try something “small” and unfamiliar outside Meta while still remaining involved as a community member. For the AI/ML community the move matters operationally and culturally. Technically, PyTorch is positioned to continue scaling with system-level improvements and product roadmaps through 2025; organizationally, Soumith argues the project is resilient without a single steward imposing taste, with skilled systems builders and maintainers ready to run releases and product direction. His departure underlines maturation of major open-source AI infrastructure—transitioning from founder-led to community- and organization-led stewardship—and signals ongoing stability for researchers and industry relying on PyTorch, even as its development priorities evolve.