Meta AI Updates: September 12, 2026

1. Helion Kernels Can Now Ship Through the Hugging Face Kernels Hub With Pre-Tuned Configs

PyTorch added Helion support to the Hugging Face Kernels project, so kernels written in Helion, the tiled DSL often described as “PyTorch with tiles”, can be packaged and distributed on the Hub and consumed without dependency wrangling. The interesting part is autotuning: hl.tile deliberately leaves tile size and memory fetch strategy unspecified and turns them into a search space, and the integration lets authors ship pre-tuned configs alongside the kernel so consumers skip the cold-start autotune entirely. Source