The card specifies ONNX opset 14 or later and describes elementwise binary addition with multidirectional, NumPy-style broadcasting.
The kernel accepts tensors a and b and returns c, the broadcasted sum. Listed logical data types are float32, float16, int32, uint32, int8, and uint8. The repository includes metadata, an operator-contract manifest, correctness and benchmark case files, and WGSL Jinja shader templates.
The documented installation target is @huggingface/kernels@0.0.1-preview.3. Consumers can load version 1 through getKernel; the card says a 40-character commit revision can instead pin exact artifact bytes. The source requires WebGPU support. It does not establish measured performance, browser or hardware coverage, production stability, or how this kernel compares with other ONNX execution paths.