In a blog post published October 20, 2026, the PyTorch team at Hugging Face described changes to the MPS backend — the acceleration path for Apple silicon Macs — in PyTorch 2.15. According to the post, the release adds FP8 support (torch.float8_e4m3fn tensors) on MPS, storing values in one byte, half the space of FP16 or BF16, and supporting scaled matrix multiplication via F.scaled_mm. The team reports mixture-of-experts models decode 5–6.5× faster than in version 2.14, citing Qwen3.6-35B-A3B improving from 9.4 to 61.3 tokens per second on an M3 Ultra. Other stated changes include specialized matrix-vector kernels for single-row linear layers, softmax over the vocabulary described as about 8× faster, additional linear-algebra operations on the GPU, and faster paths for convolutions, padding and reductions. The post states Wan2.2's video VAE decode now peaks at 8 GiB of memory instead of 270 GiB. PyTorch 2.15 requires macOS 15 or later and Python 3.11 or later. The team notes much of the work came from running models from the Hugging Face Hub and replacing MPSGraph with dedicated Metal kernels, and previews further attention and decoding work for version 2.16.
- MoE model Qwen3.6-35B-A3B decoding rose from 9.4 to 61.3 tokens/s on an M3 Ultra, a reported 5–6.5× gain over 2.14
- FP8 (float8_e4m3fn) tensors now supported on MPS at one byte per value
- Wan2.2 video VAE decode now peaks at 8 GiB instead of 270 GiB; requires macOS 15+ and Python 3.11+
What it means for you
If your team runs or trains AI models directly on Apple silicon Macs, this release makes that faster and less memory-hungry, which can let a Mac handle models that previously wouldn't fit. For everyone else, this is deep infrastructure plumbing that happens underneath the tools you actually touch — it won't change your day.
Try this
If someone on your team already runs models locally on a Mac for development, ask them to note which models currently fail or run out of memory — that list tells you whether a PyTorch 2.15 upgrade is worth testing in a sandbox before touching anything production.
Who should care
Developers and ML engineers who run or fine-tune models locally on Apple silicon Macs, or teams using transformers, diffusers or ComfyUI on Mac hardware.
Skip this if
You use AI through hosted services or finished products and don't run PyTorch on your own Mac hardware — this changes nothing for you.
Sources: Hugging Face — read the original