Google open-sources TorchTPU: run your PyTorch on TPUs with one line
A real escape hatch from Nvidia — eager-mode PyTorch and vLLM serving on TPUs near 2¢/M tokens, plus MiniMax stealth model ships and Notion agents run all day.
Copy markdown
Swap one line, and your model runs on a TPU
TorchTPU is a native PyTorch backend from Google and Meta: change `model.to('cuda')` to `model.to('tpu')` and your model runs out of the box, with eager execution as a first-class mode — the DeferAndFuse mode is 50%+ faster than strict eager. `torch.compile(model, backend='tpu')` works too, so you can debug on TPU the way you would on a GPU instead of compile-and-pray.
Will your model actually run? Mostly, yes
Google ran 4,917 Hugging Face models through it: 98% ran without crashing and 97% matched GPU numerics, across ~450 ATen ops with FSDP, DTensor and DDP supported. Know the gaps before you port — FlexAttention isn't supported yet, and RL colocation won't work out of the box, since a TPU takes only one process.
Serve LLMs on TPU with one pip install
`pip install vllm-torchtpu` gives you the same vLLM CLI and OpenAI-compatible server on Cloud TPUs, with Ironwood (v7x) the recommended chip. SemiAnalysis clocked Qwen 3.5 at ~2.01¢ per million tokens, down from 3.6¢ in September, and 0.83¢ on its agent profile — and there's a free Community TPU pool for people building on it.
The CUDA moat cracked twice in a day
TorchTPU lands barely a day after Amazon's TorchNeuron put stock PyTorch on Trainium, so the two biggest non-Nvidia accelerators now both take unmodified PyTorch. That's real leverage on your bill: the same training and inference code can shop for the cheapest silicon instead of being welded to CUDA.
MiniMax's stealth 'Space Bunny' ships as M3.1 Flash
The mystery model testers were hammering is MiniMax M3.1 Flash: a 428B-param MoE (~23B active) with a 1M-token context, tuned for agentic coding and computer use. It's $0.15 / $0.60 per million tokens in/out — roughly 70% cheaper on agent tasks than M3 — with a free rate-limited endpoint on OpenRouter and open weights promised Oct 23.
Notion's agents now run for a full day
Notion's 6th-gen agent harness lets a single task run 24+ hours, up from about five minutes, on queue-backed workers that survive restarts, with a new Auto router picking models for ~90% lower average cost. If you build on Notion's agent platform, long-running jobs just stopped timing out on you.