The post says users can load a Transformers model with device_map="tpu" and move a Diffusers pipeline with .to("tpu"); it also describes distributed training and torch.compile integration. Hugging Face says Optimum TPU is no longer required to run its Transformers and Diffusers models on TPUs.
The announcement reports that 96.5% of the tests run on TPU pass, with 34 issues reported and 19 fixed. It also gives workload-specific performance examples: at batch size 64, the company reports decoding 21.5 times faster for Qwen3-30B-A3B and 12.8 times faster for Gemma 4 26B-A4B using Hub kernels, measured on a TPU v6e-8. Those figures are not a general comparison of TPU and GPU performance or cost.