PyTorch 2.15 Makes Your Mac Run AI Models Up to Six Times Faster

Key points

• PyTorch 2.15 speeds up mixture-of-experts language models, a type of AI that routes each calculation to specialised sub-networks, by 5 to 6.5 times on Apple silicon Macs compared with version 2.14.
• A memory bug in video generation using the Wan2.2 model has been fixed: peak memory drops from 270 gigabytes to roughly 8 gigabytes.
• FP8, a compact number format that stores weights in half the space of the previous standard, is now supported on the Mac's GPU for the first time.
• A silent image-corruption bug affecting the Qwen-Image-2.1 model on Mac has been patched; image quality jumps from a broken 10.06 dB to a correct 29.95 dB.
• PyTorch 2.15 requires macOS 15 or later and Python 3.11 or later.

If you've ever tried running an AI model on a Mac and watched it crawl, PyTorch 2.15 is the update you've been waiting for. The Hugging Face PyTorch team published the details this week, and the numbers are striking.

The headline result involves mixture-of-experts models, a type of large language model where each calculation is routed to one of several specialised sub-networks rather than running through the whole model every time. On an M3 Ultra chip with compiler optimisations turned on, Qwen3.6-35B-A3B goes from 9.4 tokens per second to 61.3. LFM2.5-8B-A1B jumps from 23.6 to 123.6 tokens per second. A token is roughly three-quarters of a word, so that second figure means the model can produce readable text in near real-time on a laptop.

What actually changed under the hood?

Several things changed at once. For language models, the biggest fix is a new native Metal kernel, a small programme that runs directly on Apple's GPU chip, for grouped matrix multiplication. The old code had to pause and consult the main processor during every calculation cycle; the new code keeps everything on the GPU, which is far more efficient for this kind of maths. Softmax, a calculation that turns a list of numbers into probabilities and runs at every decoding step, is now about eight times faster on its own.

Video generation got a different kind of fix. Wan2.2's video decoder previously expanded data into a 270-gigabyte chunk in memory, far beyond what any Mac actually has, which is why it crashed for most users. Breaking the job into smaller tiles brings that peak to about 8 gigabytes.

FP8 support rounds out the major additions. One byte per value, half the storage of BF16, the format most models currently use. Apple GPU users can now load FP8 weights and multiply them on the GPU without converting to a larger format first. For anyone running a big model on a Mac with limited memory, that's a meaningful change.

What does this mean for ordinary Mac users?

When Hugging Face added one-line GGUF support for Apple silicon Macs on 22 September, the focus was on making compressed models easier to load. What's different here is that these gains come not from model compression or new hardware but from better software talking to the chip Apple already shipped.

If you run local AI tools such as ComfyUI or the transformers and diffusers libraries, upgrading to PyTorch 2.15 should improve speed and stability without any changes to your workflow. The new version requires macOS 15 (Sequoia) and Python 3.11, so check both before upgrading.

The team has been running real models from the Hugging Face Hub to find and fix weak spots rather than relying on synthetic benchmarks. That approach shows in the bug fixes: the Qwen-Image-2.1 image corruption issue was a silent failure that produced damaged outputs without any error message. Users had no way to know their edits were broken.

PyTorch 2.16 is expected to move attention calculations fully onto Metal kernels and push decoding speeds higher still. Watch whether those gains hold across M1 and M2 chips, not just the M3 Ultra used in these benchmarks, because that's where most Mac users actually are.

The update is available now as a release candidate.

Common questions

Do I need a new Mac to benefit from this?

No. The improvements work on existing Apple silicon Macs, from M1 onwards, as long as you run macOS 15 or later. The biggest speed gains appear on M3 Ultra chips, but all Apple silicon users should see some improvement.

Will my existing AI tools update automatically?

No. You need to upgrade PyTorch to version 2.15 manually. Tools like ComfyUI will then use the faster backend automatically once PyTorch is updated.

Is FP8 safe to use for everyday models?

FP8 is a lower-precision format, meaning it stores less detail per number. For most inference tasks, meaning asking a model a question rather than training it, the quality difference is negligible. It's best suited to running large models that would otherwise not fit in memory.