ONNX · 3 models available
Run a decision model with ONNX Runtime
An ONNX export loads without PyTorch. That matters for slim containers, for edge devices, and for running a decision inside a browser.
How to run it
Install the runtime
ONNX Runtime loads without PyTorch, which is usually the point. Container images get much smaller.
pip install onnxruntime huggingface_hubFetch the exported graph
ONNX repositories ship a .onnx file plus the tokeniser. Download both.
Consider the browser
ONNX Runtime Web and transformers.js run the same graph client side, which puts the decision on the user’s machine with no server call at all.
Models available in ONNX
Every entry in the directory that ships this format, read from its repository metadata on 20 September 2026. None has been benchmarked by us.
rlcd-modernbert-151m
The smallest decision model we track, at 151M parameters on a GLiClass/ModernBERT encoder. If it holds accuracy, it changes what this class of model costs to run.
laya-onnx
A ONNX build of laya, for running typed decisions on runtimes without PyTorch instead of a hosted API.
laya-onnx-fp16
A ONNX build of laya, for running typed decisions on runtimes without PyTorch instead of a hosted API.
Also browsable as a facet of the directory: ONNX builds.
Other ways to run one
GGUF
GGUF is the format that made local inference ordinary. For decision models it matters more than usual, because the whole argument for the category is cost per call.
MLX
MLX uses unified memory on M-series chips. For a model in the few-hundred-megabyte range this is the fastest local option on a Mac, and it needs no GPU of its own.
Core ML
Core ML puts the model on the Neural Engine, so the decision happens on the phone or the laptop and nothing leaves it. The strongest privacy story in the category.
Compatible servers
16 open projects that reimplement the decision API, if you want the interface rather than just the weights.
Find out what your data actually supports
We turn your historical decisions into a frozen test set, then tell you which model, which threshold, and how much of it you can safely automate.