Amazon’s Trainium chips have long been praised for their speed but criticized for their awkwardness in AI research. The issue? They excel at running finished AI models but struggle with the experimental, messy work of building new ones. To address this, AWS has launched TorchNeuron, a native PyTorch backend for its Trainium2 and Trainium3 accelerators, designed to streamline research training.
TorchNeuron entered public beta on October 19, 2026, and was open-sourced the following day during the PyTorch Conference North America. The release coincided with version 2.33.0 of the Neuron SDK, AWS’s toolkit for programming its in-house chips. The standout feature is simplicity: developers can now move a model onto Trainium with a single line of code, such as .to(‘neuron’), making the process feel like any other PyTorch job.
The update also includes several supporting upgrades, such as Neuron Observer for detecting hardware health issues, improved NKI functionality for lower-level customization, and updated vLLM plugins tuned to work with TorchNeuron. However, there’s a catch: TorchNeuron only supports Trainium2 and Trainium3, leaving out older hardware like Trainium1 and the Inferentia chip family.
AWS reports that over 30 early customers are already using TorchNeuron, spanning fields like robotics, voice AI, and large language model development. Companies such as Cartesia and Deepgram have reported substantial performance improvements, while Hugging Face has adopted Trainium support. The update is expected to open Trainium to broader research use cases, including drug discovery.