Hugging Face × TorchTPU, Claude Code 2.1.305, [Late catch] Timestral

Hugging Face published how to run Transformers, Diffusers and Kernels on TorchTPU, a PyTorch backend that treats a TPU as just another device. It replaces the older PyTorch/XLA route and Optimum TPU. Install it on a TPU VM:

pip install torch-tpu --extra-index-url https://download.pytorch.org/whl/cpu

Then load models with device_map="tpu" and move inputs with .to("tpu"). Eager mode is the default, and torch.compile is available for more speed.

Moving GPU code to TPUs used to mean writing XLA-specific code, and that barrier just got lower. In Hugging Face's own numbers, Qwen-Image-2.1 at 1024×1024 took 3.5 s compiled versus 8.0 s eagerVendor claim, and MoE decoding got up to 21.5x faster with a dedicated kernelVendor claim. Teams short on GPUs now have a cheaper path to evaluate TPUs.

The useful addition in this week's builds is /auto-mode-setup, a guided setup that teaches auto mode about your environment and works on Bedrock, Vertex, Foundry and other providers. /claude-api managed-agents-secure compares a Managed Agents setup with Anthropic's security guide without changing anything. On Pro and Max, resuming an idle conversation now warns you when it would eat a large share of the 5-hour limit.

Many teams left auto mode off because writing allow rules by hand was a chore, and guided setup makes starting easier. If you run behind an internal gateway, tool search now defaults on and retries on its own after a 400 or 422. Read the note on sandbox.allowNestedUserNamespaces in 2.1.305 before enabling it, since it can bypass the sandbox's file protections.

Timestral is Mistral's first family of time-series foundation models. It forecasts sensor readings, demand or cash flow zero-shot, meaning without per-dataset training, and returns probabilistic forecasts over 21 quantiles. It comes in four sizes (5M, 28M, 123M, 209M parameters) with a 32,768-step context, and the weights are on GitHub and Hugging Face.

Mistral says the 209M model ranks first on GIFT-Eval, TIME and fev-benchVendor claim. It reports about 10 ms per forecast for the 5M model and about 140 ms for 209M on 8 CPU threadsVendor claim, so you can try demand forecasting or anomaly detection without a GPU. The announcement says modified MIT while the repo lists Apache-2.0, so check the license before commercial use.