AWS launched TorchNeuron, a native PyTorch backend for its Trainium2 and Trainium3 accelerators, entering public beta on October 19, 2026, and open-sourcing the code on October 20, 2026, during the PyTorch Conference North America. The release aims to simplify research training on Trainium chips, which have been effective for running finished AI models but cumbersome for experimental model development. TorchNeuron shipped alongside Neuron SDK version 2.33.0 and enables developers to migrate models to Trainium hardware with a single line of code, addressing the complexity of the previous PyTorch/XLA integration that required custom programming patterns.
TorchNeuron Supports Trainium2 and Trainium3 Only
TorchNeuron supports Trainium2 and Trainium3 accelerators exclusively. Trainium1 and the Inferentia chip family are not compatible with the new backend. Developers can move a model onto Trainium with a single line such as .to('neuron'), maintaining compatibility with standard PyTorch APIs. The backend supports eager execution, meaning operations run immediately as the code calls them instead of being compiled into a larger graph first. The previous PyTorch/XLA integration required developers to insert calls like xm.step() throughout training loops and work within a compiler-heavy programming style.
Neuron SDK 2.33.0 Adds Observer Tool and vLLM Updates
Neuron SDK version 2.33.0 includes Neuron Observer, a tool for detecting hardware health issues, improved NKI functionality for lower-level customization, and updated vLLM plugins tuned to work with TorchNeuron. Trainium has found use in inference, the stage where a trained model answers real requests. Anthropic runs production inference and reinforcement learning rollouts on Trainium2 and Trainium3.
Anthropic and Hugging Face Among Early Adopters
AWS reports more than 30 early customers using TorchNeuron, spanning robotics, voice AI, and large language model development. Cartesia and Deepgram, two companies working in voice AI, report substantial performance improvements compared with their earlier configurations. Hugging Face has adopted Trainium support. The update opens Trainium to broader research use cases, with fields such as drug discovery cited as areas that could benefit from easier access to the hardware.
FAQ
What did AWS launch on October 19, 2026?
AWS launched TorchNeuron, a native PyTorch backend for its Trainium2 and Trainium3 accelerators, entering public beta on October 19, 2026, and open-sourcing the code on October 20, 2026, during the PyTorch Conference North America.
Which Trainium chips support TorchNeuron?
TorchNeuron supports only Trainium2 and Trainium3 accelerators. Trainium1 and the Inferentia chip family are not compatible with the new backend.
Who are the early customers using TorchNeuron?
AWS reports more than 30 early customers using TorchNeuron, including Anthropic, Cartesia, Deepgram, and Hugging Face, spanning robotics, voice AI, and large language model development.
