Open-weight AI modelsWed 21 Oct 2026 · 3 of 12

Hugging Face launches TorchTPU for native PyTorch support on Google Cloud TPUs

Verified · 6/6 supported1 cited · 1 min read

Hugging Face introduced TorchTPU, a native PyTorch backend for TPUs . With TorchTPU, Transformers, Diffusers and Kernels run natively on TPUs within the Hugging Face ecosystem . TPUs are Google’s AI accelerators built for matrix math and are used by Google to train and serve models such as Gemini . Google Cloud offers TPUs in configurations ranging from a single chip to pods of thousands of chips . Previously, PyTorch on TPUs required torch_xla with lazy tensors and Optimum TPU for Hugging Face models, but those are no longer needed with TorchTPU .

Why it matters. The change lets developers run Hugging Face models on Google Cloud TPUs directly, removing the need for Optimum TPU and simplifying training and inference workflows .

Earlier on Open-weight AI models

Tue 20 Oct 2026TorchNeuron public beta simplifies Hugging Face model deployment on AWS TrainiumTorchNeuron is a PyTorch plugin that adds a new backend for AWS Trainium, supporting eager mode, torch.compile, and native distributed APIs such as FSDP. Starting on October 20, 2026, TorchNeuron will be released as a public beta and can be installed with pip install torch-neuronx. Previously, Hugging Face workloads on Trainium relied on the optimum-neuron library, which required lazy tensors, static shapes, and recompilation for each graph break. TorchNeuron integrates directly into PyTorch, exposing Neuron as a native device with eager execution, PyTorch’s distributed primitives (DTensor, FSDP2), and torch.compile, allowing libraries like Transformers, Diffusers, Accelerate, PEFT and TRL to run on Trainium like any other accelerator.Fri 16 Oct 2026Meta.ai launches Muse Spark Open, a desktop‑capable multimodal AI modelmeta.ai reports that Muse Spark Open is available today as the most capable US open‑weight model to date. meta.ai reports that the model is natively multimodal with 816 billion total parameters (42 billion active) and a 1 million‑token context window. meta.ai reports that it is built for coding and agentic work, excelling at long‑horizon tasks, complex software engineering, and multi‑step reasoning. meta.ai reports that Muse Spark Open is the first model in its class that can run on a desktop using two Apple M5 Ultra Mac Studios via the MLX framework, and it is available through API platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI, with model weights downloadable from Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models. meta.ai reports that the model runs on inference engines such as SGLang, vLLM, TensorRT‑LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth, with optimizations for AMD, Apple, Intel, and NVIDIA silicon, and developers can further customize it using the PyTorch framework.Fri 16 Oct 2026Granite Switch 4.1 unifies chat, decision‑making and speech on a VRAM‑efficient endpointGranite Switch 4.1 includes Granite‑Switch‑4.1‑8B and -30B models equipped with a new Decision aLoRA for general‑purpose decision‑making. The models also contain a specialized Guardian aLoRA that outperforms Jev 1.13, Microsoft‑Decision‑1, GPT‑Luna‑Decisions, and Clef‑Flash on safety‑related decision tasks. Granite Switch 4.1 integrates Granite‑Speech‑5.0‑Turbo‑CTC, the fastest ASR model on the OpenASR Leaderboard, for transcription tasks. All three adapters plus the base model use roughly the same VRAM as the base model alone, allowing more capabilities with fewer GPUs. The cache‑efficient Activated LoRAs enable reuse of the KV cache, and combined with vLLM speed‑up and a co‑located ASR encoder the system can switch between chat, decision, and spoken requests with blazing fast speeds on a single endpoint.

Next · Health researchStudy links autoimmune rheumatic diseases to higher long COVID risk →