- The new PyTorch 2.15 speeds up AI models on Apple computers: according to the authors up to 5 to 6.5 times in selected tests.
- It requires macOS 15 and Python 3.11 or newer.
- The numbers come from the authors of the changes and cover one computer, so measure your own model.
The PyTorch team at Hugging Face published an overview of what is new in PyTorch 2.15 on 20 October 2026. In short: AI models are meant to run much faster on Apple computers and use less memory. The authors write that at the start of 2026 even small models on a Mac could stutter or suddenly slow down, because support for some operations was missing. Their work consists of running real models from the Hub and patching such gaps.
What changed
- Faster “experts” models (MoE). According to the authors they decode 5 to 6.5 times faster than in version 2.14 in selected measurements. An example from the post: the model Qwen3.6-35B-A3B goes from 9.4 to 61.3 tokens per second on an M3 Ultra computer (with torch.compile), and LFM2.5-8B-A1B from 23.6 to 123.6.
- Numbers written as FP8. This format takes half the space of BF16 and can be calculated on directly on the graphics chip.
- A faster choice of the next word. The final step where the model picks a word from the whole vocabulary (softmax) is, according to the authors, about eight times faster. This is a microbenchmark result, meaning a measurement of one small piece.
- Less memory for video. Decoding video with the Wan2.2 VAE needs a peak of about 8 GiB instead of 270 GiB.
- A fixed bug. In version 2.14 images in the VAE encoding of the Qwen-Image-2.1 model on MPS came out corrupted.
What you need
MPS now requires macOS 15 or newer, and PyTorch 2.15 requires Python 3.11 or newer. If the system is too old or the computer has no chip with Metal support (Apple’s graphics technology), the MPS availability check gives the reason. The post also previews version 2.16, where attention, a key building block of language models, is to be calculated fully on Metal.
How to read it
The numbers apply to specific models and one computer, so treat them as a direction, not a promise. Before you update the environment, check the system and Python version and run your own test on twenty examples, as in the article What fits on your own computer. I compare local tools in the text Ollama, LM Studio and llama.cpp compared, and you will find the whole subject in the Local AI hub.