Does Your Model Run on AWS Trainium? Look It Up

Hugging Face has published a lookup tool showing which popular models on its Hub run unmodified on AWS Trainium chips. The post, written by David Corvoysier and dated October 23, 2026, follows the company's introduction of TorchNeuron, a native PyTorch backend for Trainium distributed as the torch-neuronx package. The new torch-neuronx model support Space lists each model's result, the number of Trainium cores used, wall time including compilation, and peak device memory. Users can browse a curated selection or search Trending and Downloads views, and filter by versions of Transformers and the neuronx-cc compiler.

For each tested model, the Space provides the software stack, required packages and the run.py script used, so developers can reproduce the test on a Trainium2 instance. Precompiled graphs can be copied from a public cache bucket to skip the initial compilation step.

Hugging Face is explicit about the limits of a passing grade. Each test loads a model across the fewest cores able to hold its weights, then generates 32 tokens. Single-core output is compared against a CPU run, while for sharded models a small judge model checks only that the text is fluent English. According to the post, a PASS "says nothing about throughput" or longer generations.

Coverage is currently restricted to Transformers text-generation and image-text-to-text models that fit on four Trainium2 devices, or 16 cores, with diffusers and sentence transformers planned. All results sit in a public dataset, giving teams weighing Trainium a way to check compatibility and failure details before committing to it.