Anyone who has tried to move a model from Nvidia GPUs to Amazon's own AI chips knows the first question is the hardest one: will it even run? Hugging Face now has a public answer. On 23 October 2026 it published a lookup page for AWS Trainium model support that shows, model by model, whether popular checkpoints on the Hugging Face Hub run unmodified on AWS Trainium2 chips.
The tool lives in a Hugging Face Space called "Which Hub models run on AWS Trainium with torch-neuronx", run by the aws-neuron organisation. It was announced in a blog post titled "Does Your Model Run on AWS Trainium? Look It Up", written by David Corvoysier. Here is what it does, what it tests and where its limits are.
What the AWS Trainium model support page shows
The page opens on a "Selected" view, a curated list of top Hub models ranked by downloads and trending score. Two more views, "Trending" and "Downloads", follow the Hub's own rankings one task at a time, whether or not a model has been tested yet. You can also search by part of a model's name.
Each result row gives four pieces of information:
- the outcome of the test, such as a pass or a recorded error;
- how many Trainium cores the model needed;
- the wall time of the run, including compilation;
- peak device memory.
You can switch between versions of the Neuron compiler (neuronx-cc) and of the Transformers library to see how support changes from one release to the next. That matters because a model that fails on one compiler version may pass on the next.
Models that have not been tested come with a short reason. According to the blog post, the usual reason is that they were not prioritised, not that they are known to be incompatible.
Why this is useful
Trainium is Amazon Web Services' in-house AI accelerator, and AWS has been pushing it as a cheaper alternative to GPUs for training and inference. The software that connects PyTorch to Trainium is called TorchNeuron, a native PyTorch backend that Hugging Face introduced in a separate post on 19 October 2026.
The practical problem for developers has been uncertainty. Before committing time and money to a new chip, a team wants to know whether the exact model it uses already works. Until now that usually meant renting an instance and trying. The lookup turns that trial and error into a search box.
It also gives a starting recipe. Each model page lists the Neuron software stack that was tested, the packages to install and the Python script used for the run. That script loads the model through the standard Transformers loading call, splits it across chips when needed and generates text with a static cache and PyTorch's compile feature pointed at the Neuron backend. The post says no separate distributed launcher is needed, because the script starts its own processes.
To save time, users can pull precompiled graphs from a public cache bucket that aws-neuron maintains, organised by compiler version, instead of compiling from scratch.
How each model is tested
The test is deliberately simple. According to Hugging Face:
- The model is loaded on a Trainium2 (trn2) instance.
- It is split across the smallest number of Trainium cores that can hold its weights. Most models use tensor parallelism; mixture-of-experts models without a tensor-parallel plan use expert parallelism only.
- The model generates 32 tokens.
- On a single core, those tokens are compared with a CPU run of the same checkpoint. For models split across several cores, a small judge model checks that the output is fluent English.
All results come from a public Hugging Face dataset, also named torch-neuronx-model-support, which stores one Parquet file per combination of compiler and library version. Each row records the result, any error, the software versions and the script that was used. That means teams can query the data directly rather than relying on the web page.
What a pass does and does not mean
The blog post is clear about this: a pass "says nothing about throughput, or about longer generations." It tells you the model loads, runs and produces sensible output for 32 tokens. It does not tell you how fast it is, how much it costs per token or whether it holds up over a long context.
Current limits
The tracker is narrow for now. Hugging Face lists these restrictions:
- Only Transformers models are tested.
- Only two tasks are covered: text generation and image-text-to-text (vision language models).
- Only models that fit on four Trainium2 devices, or at most 16 cores, are tested. Very large models are out of scope.
- Only Trainium2 is named. The post does not mention first-generation Trainium, Trainium3 or Inferentia.
Hugging Face says it plans to add diffusers image models, sentence transformers and more Transformers tasks. The post does not describe a way for outside users to submit their own test results.
Who should look at it
The page is most useful for three groups:
- Teams already on AWS weighing Trainium against GPU instances for inference of small and mid-sized open models.
- Developers choosing an open model who want one that will run on more than one kind of hardware.
- Researchers and startups with AWS credits who want a quick way to check whether a model is worth trying on Trainium before spending them.
If you are comparing small open models to run yourself, our recent look at Recursive Flint, a free coding model that runs on a laptop covers the other end of the hardware range.
What to watch
The lookup is a small tool, but it reflects a bigger shift. Hugging Face also published a post on TorchTPU, for Google's chips, on 20 October. Both companies are working to make non-Nvidia hardware feel like a normal PyTorch target. Whether Trainium wins more users will depend on the numbers this page does not yet show: speed and cost per token. For now, it answers the first question developers ask, and it does so in public.
