PyTorch made running a model on Mac’s up to 12x faster
PyTorch made running a model on the Mac you already own up to 12x faster. PyTorch 2.13 shipped on 8 July. Its headline change is for people running models on Apple hardware: FlexAttention now runs on Apple Silicon, with up to ~12x speedups over the old attention path on sparse workloads. In the same release, on-device inference (ExecuTorch) became a built-in part of the core framework, and PyTorch now reads safetensors weight files natively — no extra library.
What it actually means
For a month, every item here has asked which model you rent. This is the counterweight: the runtime for running a model on the laptop already on your desk just got materially faster, and it's free. It matters now because the other half has arrived. Open weights worth running are dropping. Fast local runtime plus downloadable frontier weights is the first time 'own it' is a real answer to 'rent it' for a small team.
The whole month argued about which model to rent. The quieter story is that owning the stack just became a real answer.
Source: PyTorch Foundation · Tuesday 21 July 2026