Amazon plans to release its TorchNeuron software as an open-source public beta on Oct. 20, aiming to cut the engineering effort involved in trying its Trainium AI chips through familiar PyTorch workflows. Amazon announced the plan Oct. 19, with the beta scheduled for the PyTorch Conference North America in San Jose, California.
Amazon says TorchNeuron will support Trainium2 and Trainium3 through PyTorch’s existing mechanism for connecting hardware. For compatible code, the company says developers can select the device through a setting rather than rewrite the model. The software also includes tools for testing, debugging and scaling models. PyTorch is an open-source framework for developing AI models, and its governance moved to the PyTorch Foundation under the Linux Foundation in 2022.
The beta will use the Apache 2.0 license. Amazon’s technical announcement says the first release will not support dynamic or symbolic shapes and lacks support for some PyTorch operations. TorchAO and FP8 support remain in private beta. Jana van Greunen, Meta’s director of PyTorch Engineering, said the effort is intended to give developers a “consistent experience across hardware platforms.”
Software support is one part of the competition for AI workloads. Nvidia’s CUDA platform provides tools for applications running on its GPUs, while Amazon says TorchNeuron can reduce the engineering time required to test Trainium. Google introduced TorchTPU on April 7, 2026, to run existing PyTorch workloads on its chips with minimal code changes.
An Oct. 19 case study by AWS and Deepgram offers one example of the work and potential payoff. The companies said Deepgram’s tuned Trainium2 setup handled 11% to 70% more audio per dollar than its previous compute options on offline batch workloads. They said the effort involved several months of porting and performance tuning.
The public beta will let developers begin evaluating TorchNeuron on their own workloads when it opens at the conference Oct. 20.
