The site seems to be loading wasm, a model_q4.onnx, and a large tokenizer on each visitor and visit. 50MB per visit. Their bandwidth costs bust be next level.