ONNX and ort

Claude Opus 5.5, 4 Nov. 2026

NLP and argumentation

ONNX (Open Neural Network Exchange) is a portable file format for trained machine-learning models. A model trained in Python with PyTorch, for example, can be exported to ONNX. ONNX Runtime then runs it efficiently on many platforms and kinds of hardware. ort is ONNX Runtime’s Rust binding.

In Quell

The labs prototype in Python, but the game is written in Rust. Exporting a lab’s models to ONNX would let Quell run exactly the model the lab measured through ort, with no Python at runtime. This is the planned path for non-generative models such as embedding, NLI and GLiNER models. Generative models would use llama.cpp instead.

External links

• ONNX — the format.
• ONNX Runtime — the runtime.
• ort — the Rust binding.

What links here: Embeddings and cosine similarity, Quantization and llama.cpp, oAMF, Reading notes