Breaking02 / IT
Hugging Face unifies model export return format
The update brings exported models under a shared interface for sharing and execution, while changing how developers save them.

Hugging Face announced a unified model export interface in Transformers v5.23.0 for execution through ONNX Runtime, without naming a licence in the announcement.
The release notes identify ExportedModel as the common return type: an object that holds an exported model. It accepts the same arguments and returns the same output format as the original model.
According to the announcement, the same interface also supports OpenVINO and ExecuTorch. Exported models can be shared on the Hub and loaded again, with execution handled by each engine's own runtime.
The release notes flag changes that break compatibility with existing code. The export() method now returns an ExportedModel, and an ONNX program now uses the onnx_ir.Model type.
The notes also announce the removal of output_path from ONNX and OpenVINO configurations. Developers use save_pretrained to save exports and register_backend to register custom execution engines.
The release notes do not specify the minimum graphics memory (VRAM) required to run exported models.