PyTorch 2.15, now in release candidate form, adds FP8 tensor support to the Metal Performance Shaders (MPS) backend, letting Apple silicon Macs store float8 values at half the size of BF16 and run scaled matrix multiplication on the GPU. The release also targets language-model speed. The Hugging Face PyTorch team, which works on the MPS backend, says mixture-of-experts models now decode 5 to 6.5 times faster than in version 2.14. On an M3 Ultra, Qwen3.6-35B-A3B rose from 9.4 to 61.3 tokens per second, and LFM2.5-8B-A1B rose from 23.6 to 123.6 tokens per second when run with torch.compile. Dense models gain from new matrix-vector kernels for single-row linear layers, which the team measured at a 1.23 times geometric-mean speedup across 96 BF16 layer shapes, and from a softmax over the vocabulary that is about 8 times faster. The update also reduces memory use and fixes bugs in media workloads. Decoding video with Wan2.2's tiled VAE previously pushed driver memory to 270 GiB, and it now peaks at about 8 GiB. A padding change fixes an issue in which the Qwen-Image-2.1 VAE encoder silently corrupted images on MPS in 2.14; an encode-decode round trip now matches CPU at 29.95 dB PSNR, up from 10.06 dB. Other changes include complex LU and triangular solves, faster QR decomposition, faster paths for convolutions and dropout, and public MPS streams in Python. The release requires macOS 15 or later and Python 3.11 or later, and the MPS availability check now reports why acceleration is unavailable on unsupported systems. For Australian businesses and professionals, the practical effect is limited to those who run models locally on Macs. There is no direct cost or regulatory change. Teams that run open-weight models on-device may find that faster local inference reduces reliance on paid cloud APIs for some workloads, and keeping data on the device can simplify privacy assessments under the Privacy Act. Because the build is a release candidate, it is best tested in a separate environment before it is used on production tooling. If you do not run PyTorch models on a Mac, this changes nothing for you.