A new Hugging Face Space shows which Hub models run on AWS Trainium

Hugging Face has published a lookup tool that tells developers whether a given Hub model runs on AWS Trainium. According to the announcement, a new Space reports each model’s result, plus a ready-made script to run it.

A Space, a script and a public results dataset

The post, written by Hugging Face’s David Corvoysier, follows the introduction of TorchNeuron, the native PyTorch backend for Trainium, shipped as torch-neuronx. Because Transformers runs its own modeling code on Trainium as it does on any other device, many Hub models may work there. The new aws-neuron/torch-neuronx-model-support Space is meant to show which ones actually do.

The Space opens on a curated “Selected” view of top models ranked by downloads and trending score. Each row lists the result, the number of Trainium cores used for sharding, the wall time of the run including compilation, and peak device memory. “Trending” and “Downloads” views follow the Hub’s own rankings one task at a time. Users can pick the Neuron compiler (neuronx-cc) and Transformers versions to inspect. When a model is absent on purpose, the Space says why: usually it was not prioritized, and occasionally it is truly incompatible.

For each model the Space names the Neuron stack it was tested on, the packages to install and the run.py script used. On a Trainium2 (trn2) instance, such as in the AWS Neuron Deep Learning Container, running python3 run.py is enough. The script launches its own torch.distributed ranks when it shards a model, so torchrun is not needed. It loads the model with from_pretrained and generates with a static cache and torch.compile(backend="neuron").

The first run compiles the model’s graphs. To skip that step, users can copy the graphs built during testing from a public NEFF cache bucket with hf buckets sync, using the command the Space shows for the matching compiler version.

What a PASS does and does not promise

Hugging Face defines “runs” narrowly. Each test loads the model on a trn2 instance, sharded across the fewest Trainium cores that can hold its weights. Most models use tensor parallelism, while mixture-of-experts models without a tensor-parallel plan use expert parallelism only. The test then generates 32 tokens. On a single core they are compared with a CPU run of the same checkpoint; for sharded models, a small judge model checks that the output is fluent English.

The limits are explicit. Only models that fit on 4 Trainium2 devices (up to 16 cores) are tested for now, and only Transformers models for text-generation and image-text-to-text. Diffusers, sentence transformers and other tasks are planned but not dated. As the post puts it, a PASS says nothing about throughput or longer generations, so it is a compatibility check, not a performance claim.

Everything in the Space comes from the public aws-neuron/torch-neuronx-model-support dataset, with one Parquet file per pair of compiler and library versions. Each row holds the result, any error, the full stack versions and the script used, and it can be queried with pandas.

Why a lookup matters more than another backend

The tool is the practical companion to a bigger change. In the earlier TorchNeuron post, Hugging Face said the plugin entered public beta on October 20, 2026, and that it now recommends using Transformers, Diffusers, Accelerate, PEFT and TRL directly on Neuron instead of Optimum Neuron. That post also claimed that 89% of the 100 most-used text-generation architectures work out of the box, against 46% with Optimum Neuron v0.4.5. These are Hugging Face’s own figures, not independently verified.

The shift is technical as much as organizational. Optimum Neuron needed Trainium-specific model implementations, so coverage lagged each new architecture. A native PyTorch backend lets the stock modeling code run, which moves the question from “has someone ported this?” to “does it compile and produce correct output?”. Only a per-model test table answers that. The same post cited Neuron kernels giving “up to a x2.4 speedup” on vision transformers, a reminder that running and running fast are separate questions.

For teams weighing Trainium, the compile cache may matter as much as the table. Compilation time is part of the cost of trying a new model, and reusing graphs built during testing removes it for models already covered. Amazon has also been publishing Trainium-side work recently, such as its GPT-OSS-20B reasoning experiment on SageMaker HyperPod, and Hugging Face remains a hub for such releases, as with its simulation-ready 3D assets.

Who can use it today, and what is missing

The Space and dataset are public, so anyone can check a model without owning hardware. Actually running one requires access to a Trainium2 instance. The post does not give prices, regional availability or a schedule for wider task coverage, and it does not say how often the table will be refreshed beyond noting that results are keyed to compiler and library versions. Models too large for 4 Trainium2 devices are outside the test for now, so large frontier-scale checkpoints will not appear as passing.