Daily AI news, sourcedOn Oct 20, PyTorch 2.15 brought FP8 to Apple Silicon and up to 6.5x faster MoE decoding on MPS, Hugging Face said.
Where it stands
On Oct 20, PyTorch 2.15 brought FP8 to Apple Silicon and up to 6.5x faster MoE decoding on MPS, Hugging Face said.
- Official
PyTorch 2.15 brings FP8 to Apple Silicon and up to 6.5x faster MoE decoding on MPS, Hugging Face says
Hugging Face Blog published this on Oct 20, 2026 (source: huggingface.co).
Hugging Face's PyTorch team says PyTorch 2.15, now at release candidate, adds FP8 (float8_e4m3fn) tensors and scaled matrix multiplication on Apple's MPS backend. Mixture-of-experts models decode 5 to 6.5 times faster than in 2.14, with Qwen3.6-35B-A3B going from 9.4 to 61.3 tokens per second on an M3 Ultra, and softmax over large vocabularies is about 8 times faster. Wan2.2's video VAE decode now peaks at about 8 GiB of memory instead of 270 GiB. MPS now requires macOS 15 or later, and PyTorch 2.15 requires Python 3.11 or later.
Source: Hugging Face Blog
- Official
- An institution announces something that has already happened, through its own channel or a founder or executive account.
- Statement by …
- A named person's views, predictions, plans or comments, including remarks reported by media when the original is unavailable.
- Reported by …
- Only media or another third party is saying it. We name the outlet.
- Unconfirmed
- Circulating online (posts, comments, leaks). No company, named person or outlet has confirmed it.
Items older than 72 hours never go in Top stories and show their original date.