Meta AI Updates: October 29, 2026

1. PyTorch 2.15 Makes the Rewritten NCCL Backend the Default and Adds FP8 on Apple Silicon

PyTorch 2.15 shipped with 3,117 commits from 583 contributors, many of the headline features authored by Meta engineers. The rewritten NCCL backend, previewed in 2.14, now backs every nccl process group, with TORCH_DIST_USE_NCCL2=0 or the nccl-legacy backend name as rollback paths, and a flight recorder that dumps a post-mortem from every rank when a collective hangs. torch.distributed gains one-sided transports with a NIXL backend and CUDA-graph-capturable transfers, while FSDP2 can keep precision-sensitive parameters such as MoE router gates in fp32 and capture a full gradient-accumulation step in one CUDA graph. NVGEMM speeds up NVFP4 decode on Blackwell, MPS adds float8_e4m3fn, _scaled_mm, public stream APIs, and a grouped_mm path with 1.32 to 1.34x higher MoE decode throughput on an M3 Ultra. The default wheel moves to CUDA 13.2 (CUDA 13.4 wheels add experimental Rubin support), ROCm 10.0 wheels arrive, and Python 3.10 support is dropped. Source

2. Meta Grants $2 Million to Five US Universities for Research Using Its Models

Meta announced research partnerships with Caltech, The Ohio State University College of Medicine, Louisiana State University, Indiana University, and UTHealth Houston, backed by $2 million in grants and access to its latest models. Projects span medicine, physics, chemistry, biology, computer science, and mathematics. Examples include an LSU computer vision system that uses Meta’s open-weight SAM models to detect tracheostomy tube displacement in infants, an Indiana University project on multimodal AI that adapts conversational agents to people with dementia, and an Ohio State effort using Meta’s open-weight models to predict heart failure with preserved ejection fraction from routine abdominal CT scans and medical records. Source