PyTorch 2.15 brings FP8 and faster decoding to Mac MPS

Hugging Face's PyTorch team detailed MPS backend gains in PyTorch 2.15: FP8 support, MoE decoding 5–6.5x faster than 2.14, and Qwen3.6-35B-A3B rising from 9.4 to 61.3 tokens/s on an M3 Ultra. Softmax is about 8x faster, and Wan2.2 video VAE decode peak memory drops from 270 GiB to 8 GiB.

• Hugging Face adds native TorchTPU PyTorch backend2026-10-20
• AnyLanguageModel 1.0 ships for local models on Apple platforms2026-10-20
• Hugging Face adds native Trainium support via TorchNeuron2026-10-19
• Building GPT-2 KV Cache from Scratch and Measuring Performance2026-10-18
• Meta releases Muse Spark Open open-weight model2026-10-15