ort

Why we included this project

Rust developers who want to run ONNX models without switching languages get one of the more mature options in ort. It wraps Microsoft's ONNX Runtime behind a native Rust API, so the same code can accelerate inference on CPU, CUDA, Metal, or whichever execution provider fits the target, yet stays light enough to run on end-user devices. One crate covers both deployment-time inference and training or fine-tuning, which saves teams from juggling separate toolchains. Real-world users back the API's stability: Hugging Face's text embedding service and Google's Magika both run on it in production. If your stack is Rust and your models ship as ONNX, ort spares you from writing and debugging your own FFI bindings.