Hugging Face AI Updates: October 21, 2026

1. Transformers, Diffusers, and Kernels Run Natively on TPUs via TorchTPU

Hugging Face. Hugging Face announced native TPU support across Transformers, Diffusers, and Kernels through TorchTPU, a new PyTorch backend that treats TPU as a regular device instead of going through torch_xla. Loading a model takes device_map="tpu" in Transformers or pipe.to("tpu") in Diffusers, with optional torch.compile and tensor-parallel sharding across chips for larger models. The company says Optimum TPU is no longer required to run Hugging Face models with Transformers and Diffusers on TPUs. Source

2. PyTorch 2.15 Brings FP8 and Faster MoE Decoding to Apple Silicon

Hugging Face. Hugging Face’s PyTorch team detailed its MPS backend work shipping in PyTorch 2.15, including FP8 (float8_e4m3fn) support with scaled matrix multiplication on Macs. 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 vocabulary softmax runs about 8 times faster. Wan2.2’s tiled video VAE decode now peaks at about 8 GiB of memory instead of 270 GiB, and MPS now requires macOS 15 or later and Python 3.11 or later. Source

3. AnyLanguageModel Reaches 1.0 for Local and Hosted Models on Apple Platforms

Hugging Face. Hugging Face released AnyLanguageModel 1.0, a drop-in replacement for Apple’s Foundation Models framework that lets Swift apps use llama.cpp, MLX, and Core ML models or hosted providers with the same session code. Tool calling and guided generation now work with every provider, and hosted support covers OpenAI, Anthropic, Gemini, Ollama, Private Cloud Compute on OS 27, and Hugging Face Inference Providers. swift-huggingface 1.0 and swift-xet 1.0 shipped alongside it for faster Hub model downloads in Swift apps. Source